<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: RitzaCo</title>
    <description>The latest articles on DEV Community by RitzaCo (@ritzaco).</description>
    <link>https://dev.to/ritzaco</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F509329%2Ff5b1a3db-bdb1-4f41-a222-770a118a8156.jpg</url>
      <title>DEV Community: RitzaCo</title>
      <link>https://dev.to/ritzaco</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/ritzaco"/>
    <language>en</language>
    <item>
      <title>What is agent identity?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Tue, 17 Nov 2026 11:00:56 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-agent-identity-50pf</link>
      <guid>https://dev.to/speakeasy/what-is-agent-identity-50pf</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/agent-identity" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Agent identity
&lt;/h2&gt;

&lt;p&gt;Agent identity is the binding of every AI agent session to a real principal in&lt;br&gt;
the organization's directory, the person driving it, or a dedicated identity&lt;br&gt;
with a named human owner. The session authenticates through the identity&lt;br&gt;
provider, every tool call is authorized against that principal's grants, and&lt;br&gt;
every allow or deny is recorded under a name.&lt;/p&gt;

&lt;p&gt;Three parts of the definition are load-bearing:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The binding runs to the directory.&lt;/strong&gt; The principal behind an agent is typically one the organization already manages, audits, and offboards.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Authorization happens per tool call.&lt;/strong&gt; An agent connected to a &lt;a href="https://modelcontextprotocol.io" rel="noopener noreferrer"&gt;Model Context Protocol (MCP)&lt;/a&gt; server is checked again each time it reaches for &lt;code&gt;create_refund&lt;/code&gt;, rather than once at connection.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The record carries a name.&lt;/strong&gt; When security asks who did something, the answer is a person, or an agent whose owner is a person, rather than a token id.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/agent-auth" rel="noopener noreferrer"&gt;Agent auth&lt;/a&gt; is the mechanism beneath this model. Identity says which principal the session binds to; agent auth says how the agent proves that binding and receives a grant on each resource request. The correct binding depends on whether the agent is &lt;a href="https://www.speakeasy.com/resources/user-steered-agents" rel="noopener noreferrer"&gt;user-steered&lt;/a&gt; or runs as a &lt;a href="https://www.speakeasy.com/resources/background-agents" rel="noopener noreferrer"&gt;background agent&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why does a shared service account fail for AI agent identity?
&lt;/h2&gt;

&lt;p&gt;The common first rollout is a bot token in the MCP config, shared by every agent on the team. Three controls that already work for people stop working at that moment:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Attribution.&lt;/strong&gt; The audit reads as the service account. After an agent issues a refund, "the billing bot did it" is usually the only available answer, and it names no one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Least privilege.&lt;/strong&gt; The key is granted the union of every task on the team, so &lt;code&gt;get_invoice&lt;/code&gt; and &lt;code&gt;create_refund&lt;/code&gt; travel on one credential and per-tool policy has no principal to attach to.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lifecycle.&lt;/strong&gt; People leave and keys do not. No directory event ends a config file's access, because the key was never bound to a person.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are the standard non-human identity failure modes applied to a caller that decides its next action at runtime. The catalog of what breaks, including the OWASP ranking of each failure, is in &lt;a href="https://www.speakeasy.com/resources/non-human-identity" rel="noopener noreferrer"&gt;what is NHI&lt;/a&gt;. The full authorization path from an identity provider to a single tool call, with the policy grammar and the audit record, is in &lt;a href="https://www.speakeasy.com/blog/governing-agent-actions-with-enterprise-identity" rel="noopener noreferrer"&gt;From your IdP to a tool call&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does agent identity bind together in a composite identity?
&lt;/h2&gt;

&lt;p&gt;A session under agent identity resolves to a composite identity with four parts:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The principal&lt;/strong&gt;, as a person's directory identity for delegated-user authority, or a dedicated agent identity with a named human owner for service or workload authority.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The agent workload&lt;/strong&gt;, so the same person driving two different agents produces two distinguishable actors.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The client&lt;/strong&gt;, whether Claude, Cursor, or a custom harness.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The task&lt;/strong&gt;, so the session's authority is scoped to the work at hand rather than to everything the person could ever do.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The composite is what makes the audit record complete, this person, through this agent, in this client, for this task. How the four parts compose, and why each one matters on its own, is the subject of &lt;a href="https://www.speakeasy.com/resources/agent-composite-identity" rel="noopener noreferrer"&gt;agent composite identity&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/agent-identity#what-does-agent-identity-bind-together-in-a-composite-identity" rel="noopener noreferrer"&gt;View the composite identity diagram on Speakeasy&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Credentials follow the same shape. The agent never holds a raw secret for an upstream system. It presents a short-lived credential minted for the task, bound to the identity, and expiring when the task completes, which is covered in &lt;a href="https://www.speakeasy.com/resources/task-scoped-credentials" rel="noopener noreferrer"&gt;task-scoped credentials&lt;/a&gt;. The session itself authenticates over OAuth 2.1 with PKCE and dynamic client registration, so modern agent clients complete the flow without custom integration.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does agent identity differ from non-human identity (NHI)?
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/non-human-identity" rel="noopener noreferrer"&gt;Non-human identity (NHI)&lt;/a&gt; is the category of digital identities that authenticate and are authorized on behalf of software, covering service accounts, workloads, pipelines, bots, and AI agents. An agent is an NHI, and the vendors who define the category, among them Okta and Oasis, now include AI agents in it alongside service accounts and tokens.&lt;/p&gt;

&lt;p&gt;Agent identity is the practice the agent class requires within that category. A service account typically runs a job someone defined in advance. An agent chooses its next call at runtime, and that difference makes the binding more consequential. The &lt;a href="https://www.speakeasy.com/resources/agent-types" rel="noopener noreferrer"&gt;four agent working arrangements&lt;/a&gt; describe accountability and engagement, while authority is a separate dimension. An agent identity must resolve to the principal behind each action: a person for delegated-user authority, a dedicated identity with a named human owner for service or workload authority, or the applicable principal at each step of a hybrid run. &lt;a href="https://www.speakeasy.com/resources/extend-iam-to-agents" rel="noopener noreferrer"&gt;Extending IAM to agents&lt;/a&gt; covers that binding and the harder question of whose permissions a custom agent built by one person and triggered by another should use.&lt;/p&gt;

&lt;h2&gt;
  
  
  What happens to AI agent access when a person is deprovisioned?
&lt;/h2&gt;

&lt;p&gt;Under agent identity, offboarding can end agent access through the same directory event that ends everything else. Directory Sync keeps users, groups, and attributes current from the identity provider, so when a person is deprovisioned, every agent session bound to their identity loses access on the same event that disables their login. Where the binding holds, there is no separate list of agent keys to find, because the agent never held a standing key of its own.&lt;/p&gt;

&lt;p&gt;Two registers make that binding checkable rather than assumed. The &lt;a href="https://www.speakeasy.com/resources/agent-inventory" rel="noopener noreferrer"&gt;agent inventory&lt;/a&gt; records each agent, the identities it uses, and its human owner of record. That field can surface an organization agent using service or workload authority whose owner has left, so the identity can be reassigned or revoked. &lt;a href="https://www.speakeasy.com/resources/agent-discovery" rel="noopener noreferrer"&gt;Agent discovery&lt;/a&gt; finds the agents that never got an identity at all, the population described in &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;what is shadow AI&lt;/a&gt;, so they can be brought under the same binding instead of running outside it.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does the Speakeasy AI Control Plane implement AI agent identity management?
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://www.speakeasy.com/product/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt; ships the binding described above as one system, on the path agents already take to tools:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Sessions authenticate through the identity provider the organization runs&lt;/strong&gt;, whether Okta, Microsoft Entra ID, Auth0, WorkOS, Google Workspace, Ping Identity, or most SAML or OIDC providers, over OAuth 2.1 with PKCE and dynamic client registration.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The platform holds the identity, never the secret.&lt;/strong&gt; It records the agent's name, its owner, its grants, and a rule saying that an assertion from a trusted issuer is that agent. It holds no secret for the agent and cannot reissue one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Every tool call is authorized against the principal's grants.&lt;/strong&gt; Agents reach tools through the platform's &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt;, so the check fires on the call, and every allow or deny is recorded under a name from the directory.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Task-scoped credentials are minted per task.&lt;/strong&gt; Short-lived, identity-bound, and expired when the task completes, so deprovisioning a person in the identity provider ends their agents' access with them.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Together, these controls replace the shared service account with AI agent identity management the organization can actually audit. The commercial page, with the identity model and the attestation classes behind it, is &lt;a href="https://www.speakeasy.com/product/agent-identity" rel="noopener noreferrer"&gt;agent identity&lt;/a&gt;. To see a session resolve to a name from your own directory, &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is agent identity?
&lt;/h3&gt;

&lt;p&gt;Agent identity is the binding of every AI agent session to a real principal in the organization's directory, the person driving it, or a dedicated identity with a named human owner. The session authenticates through the identity provider, every tool call is authorized against that principal's grants, and every allow or deny is recorded under a name. It replaces the shared service account, which gives every agent one key and leaves no trail.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why is a shared service account a problem for AI agent identity?
&lt;/h3&gt;

&lt;p&gt;A shared service account is one key presented by every agent on the team, so three controls fail at once. The audit attributes every action to the account, and it is usually impossible to say which person issued a refund. The key carries the union of every task on the team, so read-only and destructive tools travel on one credential. And when a person leaves, nothing in the directory ends the key's access, because the key was never bound to them.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is a composite identity for an AI agent?
&lt;/h3&gt;

&lt;p&gt;A composite identity binds four things into one session, the person from the directory, the agent workload, the client it runs in, and the task at hand. Authorization then has a complete answer to who is acting, this person, through this agent, in this client, for this task. Each tool call is evaluated against the person's grants rather than against a standing machine credential.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does agent identity differ from non-human identity (NHI)?
&lt;/h3&gt;

&lt;p&gt;Non-human identity (NHI) is the category of identities for software, such as service accounts, workloads, pipelines, bots, and AI agents. Agent identity is the practice of binding the agent class of NHI to a directory principal. An agent chooses its next call at runtime and usually acts on a person's behalf, so its identity has to resolve to that person, or to a dedicated identity with a named human owner, rather than to a standing machine credential that answers for the whole team.&lt;/p&gt;

&lt;h3&gt;
  
  
  What happens to an agent's access when its owner is deprovisioned?
&lt;/h3&gt;

&lt;p&gt;Under agent identity, delegated access ends with the person's directory account. Directory Sync keeps users, groups, and attributes current from the identity provider, so deprovisioning ends sessions bound to that identity. An organization agent using service or workload authority is reassigned or revoked through its owner of record rather than left running under a departed sponsor.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are task-scoped credentials?
&lt;/h3&gt;

&lt;p&gt;Task-scoped credentials are short-lived credentials minted for one task and bound to the identity behind the session. The agent never sees the raw secret for an upstream system. It presents the scoped credential, which expires when the task completes, so there is no long-lived token to leak or reuse. Combined with per-call authorization, the credential is only ever as broad as the principal's grants and only ever as long as the work.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does the Speakeasy AI Control Plane implement AI agent identity management?
&lt;/h3&gt;

&lt;p&gt;Sessions authenticate through the identity provider the organization already runs, whether Okta, Microsoft Entra ID, Auth0, WorkOS, Google Workspace, Ping Identity, or most SAML or OIDC providers, using OAuth 2.1 with PKCE and dynamic client registration. The platform holds the agent's name, its owner, its grants, and a rule that an assertion from a trusted issuer is that agent. It holds no secret for the agent and cannot reissue one. Every tool call is authorized against the principal's grants, and every allow or deny is recorded under a name.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What is MCP governance?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Tue, 17 Nov 2026 09:54:03 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-mcp-governance-42i2</link>
      <guid>https://dev.to/speakeasy/what-is-mcp-governance-42i2</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/mcp-governance" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  MCP governance
&lt;/h2&gt;

&lt;p&gt;MCP governance is the process of managing the Model Context Protocol (MCP)&lt;br&gt;
servers an organization's AI agents connect to as owned assets with a&lt;br&gt;
lifecycle. Each server is discovered, reviewed, promoted into an approved&lt;br&gt;
registry, distributed by role, monitored through a tool-call audit log, and&lt;br&gt;
updated or retired, and every stage writes to one MCP inventory.&lt;/p&gt;

&lt;p&gt;Agents connected to the tools a company uses are infinitely more useful than agents without any access. Consequently, MCP sprawl and &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP&lt;/a&gt; are becoming problems at every company with serious AI use. Adoption has been outpacing governance. And yet, as AI use has grown, and agents have become increasingly independent, it's become clear that the MCP servers agents use need to be managed the same way as any software product the company licenses.&lt;/p&gt;

&lt;p&gt;IT and security teams are now catching up and putting together comprehensive MCP governance programs that treat each server as an asset with a lifecycle. Every server, sanctioned or not, gets an owner, a recorded approval decision, distribution to the users (agents) that need it, a log of what it is doing, and a point at which it is re-reviewed or retired. Collectively, the managed servers form the company's &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;MCP inventory&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  The six stages of the MCP server lifecycle
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Discover.&lt;/strong&gt; Find the MCP servers agents already call, including &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP&lt;/a&gt; servers nobody approved, and record each one in an &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;MCP inventory&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review and approve.&lt;/strong&gt; Give each server an owner, gather evidence about what it can do and what it asks for, and record a decision.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Promote.&lt;/strong&gt; Move approved servers into the &lt;a href="https://www.speakeasy.com/resources/mcp-registry" rel="noopener noreferrer"&gt;approved registry&lt;/a&gt;, with authentication and permitted tools configured once for everyone.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Distribute.&lt;/strong&gt; Provision servers to the roles that need them, from identity provider groups, instead of per-laptop config edits.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Monitor.&lt;/strong&gt; Record every tool call in an audit log tied to the user, agent, server, and tool.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Update or retire.&lt;/strong&gt; Re-review servers whose scopes or tools change, and revoke servers that go unused, lose their owner, or fail review.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What happens at each stage of the MCP server lifecycle?
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/mcp-governance#what-happens-at-each-stage-of-the-mcp-server-lifecycle" rel="noopener noreferrer"&gt;View the MCP server lifecycle diagram on Speakeasy&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Each stage has an output, an owner, and a common way it fails. The stages below are listed in the order a new server moves through them.&lt;/p&gt;

&lt;h3&gt;
  
  
  Discover: find the servers agents already use
&lt;/h3&gt;

&lt;p&gt;Discovery compares the servers agents are observed in production against the servers the organization has approved, and surfaces the gaps between the two. Non-approved servers are considered shadow MCP, and the team that maintains the inventory needs to evaluate each server on an individual basis.&lt;/p&gt;

&lt;p&gt;One of the challenges of discovery is that, by definition, these servers fall outside the approved path. That can make detection difficult. Some companies rely on network telemetry to catch remote MCP traffic. A better option is agent-side hooks that check which MCP tools are being used within every agent session.&lt;/p&gt;

&lt;p&gt;Whatever the mechanism, the activity needs to be added to an immutable record in the &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;MCP inventory&lt;/a&gt; for every server, sanctioned or not, with its URL or package, transport, credential type, users, and last-seen time. A common point of failure is treating discovery as a one-time scan, rather than ongoing monitoring.&lt;/p&gt;

&lt;h3&gt;
  
  
  Review and approve: assign an owner and record a decision
&lt;/h3&gt;

&lt;p&gt;Every discovered server needs someone to answer three questions: who operates it, how it is authenticated (OAuth scopes, secrets, headers), and what capabilities it enables, especially the ones that write or delete. The reviewer records a decision and its scope. An approval might cover one requester, one team, or the whole organization.&lt;/p&gt;

&lt;p&gt;The review should sit on the server's inventory record, with the evidence and rationale attached, so the next reviewer can see why a server was approved or denied. The common failure is review in chat threads or tickets that nobody can find six months later, which forces the same server through review twice.&lt;/p&gt;

&lt;h3&gt;
  
  
  Promote: move approved servers into the official registry
&lt;/h3&gt;

&lt;p&gt;Promotion turns an approved server into a supported one. The server gets an entry in the organization's &lt;a href="https://www.speakeasy.com/resources/mcp-registry" rel="noopener noreferrer"&gt;approved MCP registry&lt;/a&gt;, its authentication is configured once (OAuth on the user's own identity, or a managed credential stored centrally), and its tool list is curated so that destructive tools can be excluded or granted separately. If the server should only be reachable from inside the company network, this is where that exposure is set, as described in &lt;a href="https://www.speakeasy.com/resources/private-mcp-gateway" rel="noopener noreferrer"&gt;private MCP gateway&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Promotion is also how shadow MCP gets resolved without friction. If discovery shows that dozens of engineers have each wired up the same vendor server with their own personal access token, promoting that server replaces every one of those unmanaged credentials with one governed entry. The common failure is an approved registry that lags behind demand, which sends people back to wiring servers up themselves.&lt;/p&gt;

&lt;h3&gt;
  
  
  Distribute: provision servers by role
&lt;/h3&gt;

&lt;p&gt;Distribution decides who receives each promoted server and delivers it without a manual step per person. Roles map from identity provider groups to servers and tools, and a new hire in the right group gets the right servers on day one. Delivery happens through the gateway or a managed client configuration, rather than someone sending an &lt;code&gt;mcp.json&lt;/code&gt; snippet in a message.&lt;/p&gt;

&lt;p&gt;Role-based distribution depends on the identity provider's groups being accurate. If the groups do not reflect who does what, the distribution inherits the error, so the first role mapping usually turns into a review of the groups. The common failure is admins provisioning servers one request at a time, which platform teams we talk to name as their most acute problem with MCP today.&lt;/p&gt;

&lt;h3&gt;
  
  
  Monitor: audit every tool call
&lt;/h3&gt;

&lt;p&gt;Monitoring records every tool call with the user, the agent or client, the server, the tool, the decision the gateway made, and the time. That audit log answers the security team's questions after an incident, and it also answers the governance questions: which servers are used, by whom, and how often.&lt;/p&gt;

&lt;p&gt;Monitoring also feeds the loop. A server nobody has called in months is a retirement candidate. A new server appearing in client telemetry is a discovery finding. A spike in denied calls to one server is a request for access that nobody has filed yet.&lt;/p&gt;

&lt;h3&gt;
  
  
  Update or retire: re-review what changes, remove what is unused
&lt;/h3&gt;

&lt;p&gt;MCP servers change after approval. A vendor adds tools, requests broader OAuth scopes, or publishes an advisory. Re-review compares what a server asks for now with what it asked for when it was approved, and flags the difference for a new decision. Checking published evidence has a limit, because a server can change its behavior without changing its interface, so re-review narrows the risk rather than removing it.&lt;/p&gt;

&lt;p&gt;Retirement removes a server when it goes unused for the review period, loses its owner, is replaced, or fails re-review. Retiring a server means revoking its credential and removing it from every role. Deleting the registry entry alone leaves the credential live and the connection easy to restore.&lt;/p&gt;

&lt;h2&gt;
  
  
  What makes MCP governance a closed loop?
&lt;/h2&gt;

&lt;p&gt;A closed loop means the output of each stage is the input to another, with no step that depends on someone remembering to do it. In practice, three connections close it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Blocked calls become review requests.&lt;/strong&gt; When an agent calls an unapproved server and the call is blocked, the user gets a way to request access from the same place, and the request lands on the server's inventory record for review.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Approvals become distribution.&lt;/strong&gt; A server approved for a team is provisioned to that team's role without a separate ticket to the platform team.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Usage becomes retirement and rediscovery.&lt;/strong&gt; The audit log sets each server's last-used date, and client telemetry keeps surfacing servers that are not in the registry.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The measure of the loop is the time between someone wanting a server and having it, provisioned by role and working. Platform teams we talk to name the same gap repeatedly: no inventory, no approval process, and no golden path, so every new server is a manual request. A loop that turns that request around quickly removes most of the reason to wire servers up outside the approved path. A loop that takes weeks produces shadow MCP.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why do most MCP gateways skip discovery?
&lt;/h2&gt;

&lt;p&gt;Most MCP gateways start at the approved registry. An administrator adds servers, the gateway enforces who can reach them, and the gateway logs what passes through it. That design assumes the organization already knows which servers exist, which is the assumption discovery exists to test.&lt;/p&gt;

&lt;p&gt;A gateway only sees its own traffic, so it cannot see a server that bypasses it. Discovery needs a sensor where agents run (hooks inside the coding agent, a device agent pushed through mobile device management, or a scan of client configuration files) or on the network. Few gateways ship that sensor as part of the same product, and the ones that do differ in which clients they cover.&lt;/p&gt;

&lt;p&gt;For MCP governance, this means checking whether a gateway covers the first stage, and if it does not, deciding what will. A registry and gateway without discovery govern the servers that were going to be well-behaved anyway.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who owns MCP governance?
&lt;/h2&gt;

&lt;p&gt;Ownership splits across three groups, and the split works best when it is written down:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Role&lt;/th&gt;
&lt;th&gt;Owns&lt;/th&gt;
&lt;th&gt;Typical team&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Program owner&lt;/td&gt;
&lt;td&gt;The inventory, the approved registry, distribution by role, and the loop's turnaround time&lt;/td&gt;
&lt;td&gt;Platform engineering, IT, or AI enablement&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Review authority&lt;/td&gt;
&lt;td&gt;The approval criteria, the decision on high-risk servers, and enforcement policy&lt;/td&gt;
&lt;td&gt;Security&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Server owner&lt;/td&gt;
&lt;td&gt;One server's configuration, credential, re-review, and retirement&lt;/td&gt;
&lt;td&gt;The team that requested or built it&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The program owner needs to be the platform or IT team, because the loop is mostly a provisioning and inventory problem. Security teams set the review criteria and make the calls on servers with write access to sensitive systems. When security owns the whole program, the loop tends to optimize for denial, and the turnaround time that keeps shadow MCP down stops being anyone's metric. At enterprise size each stage becomes a queue with its own owner, and the &lt;a href="https://www.speakeasy.com/resources/enterprise-mcp-gateway" rel="noopener noreferrer"&gt;enterprise MCP gateway&lt;/a&gt; page works through what a slow promotion queue costs.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you start an MCP governance program?
&lt;/h2&gt;

&lt;p&gt;Start in the order of the lifecycle, and turn on enforcement last:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Run discovery in observe-only mode.&lt;/strong&gt; Record every server agents call without blocking anything, long enough to see the servers teams use weekly as well as daily.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Triage by exposure.&lt;/strong&gt; Sort the inventory by number of users and by whether the server can write, and review from the top.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Promote the most-used servers first.&lt;/strong&gt; Replacing the servers most people already use removes the most unmanaged credentials per review.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Map roles from the identity provider.&lt;/strong&gt; Start with a few broad roles and narrow them as usage data comes in.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Turn on blocking with a request path.&lt;/strong&gt; Block unapproved servers only once a blocked user can request access from the block itself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Set a re-review and retirement schedule.&lt;/strong&gt; Decide the idle period after which a server is reviewed for retirement, and who reviews it.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Which tools support MCP governance?
&lt;/h2&gt;

&lt;p&gt;The tools below are listed by how much of the lifecycle they cover. Each entry describes only what the vendor documents.&lt;/p&gt;

&lt;h3&gt;
  
  
  Speakeasy
&lt;/h3&gt;

&lt;p&gt;The &lt;a href="https://www.speakeasy.com/product/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt; runs the full MCP governance loop on one &lt;a href="https://www.speakeasy.com/product/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Discover.&lt;/strong&gt; &lt;a href="https://www.speakeasy.com/blog/release-shadow-mcp-detection" rel="noopener noreferrer"&gt;Shadow MCP detection&lt;/a&gt; classifies every MCP tool call in an instrumented agent session and flags calls to servers outside the platform, including third-party URLs and local &lt;code&gt;stdio&lt;/code&gt; servers. Findings land in a &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/secure/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP inventory&lt;/a&gt; with last-called and last-seen times, and a flagging policy builds that inventory before any blocking starts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review and approve.&lt;/strong&gt; Each discovered server has an &lt;a href="https://www.speakeasy.com/blog/release-mcp-risk-approval" rel="noopener noreferrer"&gt;access review&lt;/a&gt; with automatically gathered evidence on its publisher, requested scopes, and tools. A blocked user gets a request link, and a reviewer approves or denies with a scope of one person, a team, or a project, with the rationale recorded.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Promote.&lt;/strong&gt; Servers enter the gateway by generation from an OpenAPI document, by adding a vendor server by URL, or by importing from a &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/catalog" rel="noopener noreferrer"&gt;catalog&lt;/a&gt; powered by the official MCP Registry, with authentication and tool curation configured per server.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Distribute.&lt;/strong&gt; &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/org-admin/roles-and-permissions" rel="noopener noreferrer"&gt;Roles and permissions&lt;/a&gt; grant access down to individual servers and tools, and roles can be assigned from the identity provider through directory sync. &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/plugins" rel="noopener noreferrer"&gt;Plugins&lt;/a&gt; bundle servers and skills and assign them to roles.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Monitor.&lt;/strong&gt; &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/observe/tool-logs" rel="noopener noreferrer"&gt;Tool logs&lt;/a&gt; record every call on the gateway, including shadow MCP calls.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Update or retire.&lt;/strong&gt; A daily check re-gathers evidence for approved servers and flags changes in OAuth scopes, credentials, or advisories for a new decision.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Two limits apply. Discovery depends on agent sessions being instrumented, so a client without the device agent or hooks produces no findings. Plugin role assignments gate delivery only through the device agent today, so a plugin installed from a client's marketplace is available to anyone with access to that marketplace.&lt;/p&gt;

&lt;h3&gt;
  
  
  Open-source gateways
&lt;/h3&gt;

&lt;p&gt;Microsoft MCP Gateway is an open-source reverse proxy and management layer for MCP servers running on Kubernetes, with Azure identity and monitoring integrations. IBM ContextForge is an open-source gateway and registry that federates MCP servers and REST services behind one endpoint. Both start from servers an administrator registers. ContextForge's "auto-discovery" refers to gateways finding each other on a network, which is a different meaning from discovering shadow MCP.&lt;/p&gt;

&lt;h3&gt;
  
  
  Commercial vendors
&lt;/h3&gt;

&lt;p&gt;Other commercial vendors building MCP gateways, registries, or MCP discovery include Runlayer, Mintive (formerly MintMCP), Kong, Docker MCP Gateway, Obot, Lasso, and Bifrost.&lt;/p&gt;

&lt;p&gt;If you're setting up the approval and provisioning loop for MCP servers your teams already use, we'd be interested to hear where it breaks down for you. &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;Talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is MCP governance?
&lt;/h3&gt;

&lt;p&gt;MCP governance is the process of managing the Model Context Protocol (MCP) servers an organization's AI agents connect to as owned assets with a lifecycle. Each server is discovered, reviewed, promoted into an approved registry, distributed by role, monitored through a tool-call audit log, and updated or retired, and every stage writes to one MCP inventory. The goal is a repeatable process that runs every week, rather than a cleanup project that is stale a month after it finishes.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are the stages of the MCP server lifecycle?
&lt;/h3&gt;

&lt;p&gt;The MCP server lifecycle has six stages. Discover finds the MCP servers agents already use, including shadow MCP. Review and approve assigns an owner and records a decision. Promote adds approved servers to the approved registry. Distribute provisions them to the roles that need them. Monitor records every tool call in an audit log. Update or retire re-reviews servers that change and removes ones that are no longer needed or no longer safe.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why does MCP governance start with inventory rather than security?
&lt;/h3&gt;

&lt;p&gt;Because an organization cannot secure an MCP server it has not found. Allowlists, data loss prevention, and tool-level permissions apply only to traffic that passes through the point enforcing them, and an MCP server wired into a local mcp.json file never touches that point. Discovery and an MCP inventory come first so that every later control has a complete list of servers to act on, and a review workflow comes next so that blocking an unapproved server gives the user a path to an approved one.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is MCP governance different from MCP security?
&lt;/h3&gt;

&lt;p&gt;MCP security is the set of runtime controls on an MCP connection: authentication, tool allowlists, logging, and inspection at the tool boundary. MCP governance is the process that decides which servers those controls apply to, who owns each one, which roles receive it, and when it is retired. Security controls without governance are enforced against an incomplete list of servers, and governance without security controls is a register nobody enforces.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is MCP governance different from agent governance and AI governance?
&lt;/h3&gt;

&lt;p&gt;AI governance covers how an organization controls what AI tools may see, do, and decide. Agent governance narrows that to the agents themselves: which agents exist, what identity they act under, and what each may reach. MCP governance narrows further to the MCP servers agents connect to, managing each server through a lifecycle from discovery to retirement. The three nest inside each other, and an MCP governance process feeds the tool side of an agent governance program.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is MCP lifecycle management the same as the MCP protocol lifecycle?
&lt;/h3&gt;

&lt;p&gt;No. The MCP specification defines a connection lifecycle with three phases, initialization, operation, and shutdown, which covers a single session between a client and a server. MCP lifecycle management, the core of MCP governance, covers an MCP server as an organizational asset over months: discovery, approval, promotion to an approved registry, distribution by role, monitoring, and retirement.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is MCP server management?
&lt;/h3&gt;

&lt;p&gt;MCP server management usually means the operational work of running MCP servers: adding them, configuring authentication, curating their tools, and deciding who can connect. MCP governance is the process that work sits inside, covering how a server is found, approved, promoted, distributed by role, monitored, and retired, with every decision recorded in an MCP inventory.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does shadow MCP fit into MCP governance?
&lt;/h3&gt;

&lt;p&gt;Shadow MCP is an MCP server employees connect to without approval, and finding it is the discover stage of MCP governance. Each shadow MCP finding becomes an MCP inventory record with an owner and a disposition. The reviewer then approves it, promotes a reviewed alternative into the approved registry, or blocks it and rotates the credential it used.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is an approved MCP registry?
&lt;/h3&gt;

&lt;p&gt;An approved MCP registry is the organization's catalog of MCP servers that have passed review, with an owner, a configured authentication method, and a set of permitted tools for each. In MCP governance, promotion is the step that moves a reviewed server into the approved registry. Pairing the registry with an MCP gateway makes approval binding, because the gateway enforces which servers and tools each caller may use.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do you distribute MCP servers by role?
&lt;/h3&gt;

&lt;p&gt;Map roles from the identity provider's groups to the MCP servers and tools each role needs, then provision servers to users through the gateway or a managed client configuration instead of per-laptop mcp.json edits. Role-based distribution depends on accurate groups. If the identity provider's groups do not reflect who does what, MCP governance inherits that error, so the first role mapping is usually a review of the groups themselves.&lt;/p&gt;

&lt;h3&gt;
  
  
  What should an MCP audit log record?
&lt;/h3&gt;

&lt;p&gt;An MCP audit log should record, for every tool call, the user, the agent or client, the MCP server, the tool, the decision the gateway made, and the time. For MCP governance, the audit log also supplies the last-used date for each server, which is the signal that decides when a server should be re-reviewed or retired.&lt;/p&gt;

&lt;h3&gt;
  
  
  When should an MCP server be retired?
&lt;/h3&gt;

&lt;p&gt;Retire an MCP server when nobody has called it for a review period the organization sets, when its owner leaves without a successor, when a vendor replaces it, or when re-review finds that its OAuth scopes, credentials, or published advisories have changed in a way the reviewer will not accept. Retirement should revoke the credential and remove the server from every role, not only delete the registry entry.&lt;/p&gt;

&lt;h3&gt;
  
  
  Do MCP gateways handle shadow MCP discovery?
&lt;/h3&gt;

&lt;p&gt;Most MCP gateways start at the approved registry and govern only the traffic routed through them, so they assume the organization already knows which MCP servers exist. Few gateways also discover the servers employees have wired up on their own, which requires a signal from the endpoint or the network rather than from the gateway's own traffic. MCP governance needs both halves, whether from one product or two.&lt;/p&gt;

&lt;h3&gt;
  
  
  Which tools support MCP governance?
&lt;/h3&gt;

&lt;p&gt;Speakeasy covers the full MCP governance loop on one platform, from shadow MCP discovery through an approved registry, role-based access, and a tool-call audit log. Open-source gateways such as Microsoft MCP Gateway and IBM ContextForge cover registration and routing for servers an organization already knows about. Commercial vendors in the space include Runlayer, Mintive, Kong, Docker MCP Gateway, Obot, Lasso, and Bifrost.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What is an MCP inventory?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Mon, 16 Nov 2026 09:12:34 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-an-mcp-inventory-1e2d</link>
      <guid>https://dev.to/speakeasy/what-is-an-mcp-inventory-1e2d</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  MCP inventory
&lt;/h2&gt;

&lt;p&gt;An MCP inventory is a continuously updated record of every Model Context&lt;br&gt;
Protocol server in an organization: the tools each server exposes, the&lt;br&gt;
credentials it authenticates with, who owns it, which agents may call it, and&lt;br&gt;
whether it is sanctioned. It records servers and tools. An agent inventory&lt;br&gt;
records the agents that call them.&lt;/p&gt;

&lt;p&gt;An endpoint scan of an engineer's laptop may list Cursor on the software register, while the same machine has a GitHub Model Context Protocol (MCP) server configured in &lt;code&gt;~/.cursor/mcp.json&lt;/code&gt;. That server can authenticate with a personal access token that reads and writes every repository the engineer can. The software register records the client, while an MCP inventory records the server, its tools, and the credential behind its access.&lt;/p&gt;

&lt;p&gt;The public &lt;a href="https://www.speakeasy.com/resources/mcp-registry" rel="noopener noreferrer"&gt;MCP Registry&lt;/a&gt; serves a different purpose. It shows which servers are available to install, while an MCP inventory records server identity, connected agents, exposed tools, authentication state, and sanctioned status in one organization. Security teams need that record to govern the servers and tools their agents call.&lt;/p&gt;

&lt;p&gt;This guide focuses on servers and tools. The &lt;a href="https://www.speakeasy.com/resources/agent-inventory" rel="noopener noreferrer"&gt;agent inventory&lt;/a&gt; covers the agents that call them.&lt;/p&gt;

&lt;p&gt;Together, the two records make MCP security policies enforceable: a policy needs to know which servers exist, which tools they expose, and which agents can call them.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why can't a CMDB or the public MCP Registry serve as an MCP inventory?
&lt;/h2&gt;

&lt;p&gt;A &lt;a href="https://www.speakeasy.com/resources/agent-discovery" rel="noopener noreferrer"&gt;configuration management database (CMDB)&lt;/a&gt; and software asset inventory track installed software. An MCP server can be a few lines of JSON in a dotfile or a remote URL. A local stdio server tells an agent which package to execute, while a remote server appears as a runtime connection. Asset scans capture the editor. The servers, tools, and credentials it reaches remain outside the record. Our guide on &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;shadow AI&lt;/a&gt; covers the wider detection problem.&lt;/p&gt;

&lt;p&gt;The public MCP Registry is a catalog that vendors and open-source maintainers publish to so clients and aggregators can discover servers. An enterprise MCP inventory covers the servers that are actually running in one organization, including:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Local stdio servers configured in dotfiles such as &lt;code&gt;~/.cursor/mcp.json&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Remote servers reached by URL, hosted by SaaS vendors&lt;/li&gt;
&lt;li&gt;Internal servers the organization built or generated from its own APIs&lt;/li&gt;
&lt;li&gt;Hosted connectors bundled inside SaaS agent platforms&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The inventory records which candidates were approved, what each one authenticates with, and whether agents are actually calling it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What belongs in an MCP inventory record?
&lt;/h2&gt;

&lt;p&gt;Each MCP inventory record needs the fields a security team will query when a credential leaks or a tool becomes destructive:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Server identity.&lt;/strong&gt; The URL for a remote server, or the package specifier and version for a local one. This identifier ties the record together across renames and vendor migrations.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Transport.&lt;/strong&gt; A stdio server runs as a child process on the endpoint, while a remote server is an HTTP endpoint elsewhere. Transport determines where the credential lives and which controls can observe the traffic.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tools and their disposition.&lt;/strong&gt; Every tool the server exposes, including whether it only reads or can create, modify, or delete data. A server whose tools are all read-only needs a different review from one exposing &lt;code&gt;create_refund&lt;/code&gt;. Tool lists can change over time, so this field needs a live feed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Credential and auth method.&lt;/strong&gt; OAuth on the caller's own identity, a managed API key, or a personal access token in a dotfile. The record should name the identity the server presents upstream. A shared credential makes every caller appear as the same actor, the failure covered in &lt;a href="https://www.speakeasy.com/resources/non-human-identity" rel="noopener noreferrer"&gt;what is NHI&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Human owner of record.&lt;/strong&gt; The person who answers for the server, renews its review, and retires it. An owner field keeps review and retirement work assigned when teams change.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Connected agents.&lt;/strong&gt; Which agents and clients are allowed to call the server. This field links to the &lt;a href="https://www.speakeasy.com/resources/agent-inventory" rel="noopener noreferrer"&gt;agent inventory&lt;/a&gt;, and each agent record links back.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sanctioned status.&lt;/strong&gt; Whether the organization approved the server. Approved servers and discovered ones sit in the same register with different dispositions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Last seen.&lt;/strong&gt; When the server last served a call. A dormant server is a revocation candidate, while an active server without approval needs investigation.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  MCP inventory vs agent inventory vs shadow MCP: what's the difference?
&lt;/h2&gt;

&lt;p&gt;An MCP inventory answers &lt;strong&gt;which servers expose a destructive tool and what each one authenticates with.&lt;/strong&gt; An &lt;a href="https://www.speakeasy.com/resources/agent-inventory" rel="noopener noreferrer"&gt;agent inventory&lt;/a&gt; answers which agents can touch billing and who owns each one. The records link to each other: each server record identifies the agents that can call it, and each agent record identifies the servers it reaches.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/mcp-inventory#mcp-inventory-vs-agent-inventory-vs-shadow-mcp-whats-the-difference" rel="noopener noreferrer"&gt;View the MCP inventory relationship diagram on Speakeasy&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The diagram shows why the records need to link. The agent inventory answers who is acting. The MCP inventory answers what it can call and how that server authenticates. Together, they show the path from an agent to the systems and data it can reach.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;Shadow MCP&lt;/a&gt; means unsanctioned servers operating outside formal security governance. Discovery surfaces those servers one finding at a time. The MCP inventory turns each finding into a record with an owner and a disposition, alongside every approved server. &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;What is shadow AI&lt;/a&gt; explains endpoint discovery, &lt;a href="https://www.speakeasy.com/blog/release-mcp-risk-approval" rel="noopener noreferrer"&gt;MCP risk approval&lt;/a&gt; shows how a finding becomes a scoped decision, and &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;what is an MCP gateway&lt;/a&gt; explains enforcement in front of the servers.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does an MCP inventory stay current?
&lt;/h2&gt;

&lt;p&gt;A static MCP inventory ages quickly. A spreadsheet compiled for an audit is accurate on the day it is finished and describes last quarter by the time anyone reads it. Because wiring up a new server is a config-file edit, a register maintained by survey decays in weeks. Staying current requires two live feeds:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Discovery on the endpoint.&lt;/strong&gt; &lt;a href="https://www.speakeasy.com/resources/ai-agent-hooks" rel="noopener noreferrer"&gt;Agent hooks&lt;/a&gt; fire inside coding agents on every prompt and tool call. They can report the MCP servers a session is actually wired to, including servers that never passed review. The &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;shadow AI reference&lt;/a&gt; describes this architecture end to end.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The sanctioned path.&lt;/strong&gt; When agents reach tools through a gateway, every session and tool call passes a point that can record the server, tool, and caller. Traffic then keeps the sanctioned-server records and last-seen field current.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How does an MCP inventory support MCP security?
&lt;/h2&gt;

&lt;p&gt;An MCP inventory makes an allow-or-block policy enforceable. A policy can block a call to a server with no record or a record marked unsanctioned. It also supports least privilege at tool granularity, allowing a policy to grant &lt;code&gt;get_invoice&lt;/code&gt; without granting &lt;code&gt;create_refund&lt;/code&gt;, one of the six controls in &lt;a href="https://www.speakeasy.com/resources/mcp-security" rel="noopener noreferrer"&gt;MCP security&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;When a server is compromised or &lt;a href="https://www.speakeasy.com/resources/mcp-tool-poisoning" rel="noopener noreferrer"&gt;rug-pulled&lt;/a&gt;, its record is where the organization revokes access. When an assessor asks which servers agents can reach and what each authenticates with, the inventory provides the evidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does the Speakeasy AI Control Plane keep the MCP inventory live?
&lt;/h2&gt;

&lt;p&gt;An organization can assemble this architecture from its existing systems. The &lt;a href="https://www.speakeasy.com/product/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt; provides it as one system and maintains the MCP inventory as servers are added and called:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The gateway is the register of sanctioned servers.&lt;/strong&gt; Servers enter the &lt;a href="https://www.speakeasy.com/product/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt; in three ways: generation from the OpenAPI documents the organization already has, adding an existing vendor server &lt;a href="https://www.speakeasy.com/blog/release-remote-mcp" rel="noopener noreferrer"&gt;remote by URL&lt;/a&gt;, or import from a &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/catalog" rel="noopener noreferrer"&gt;catalog&lt;/a&gt; powered by the official MCP Registry. The catalog identifies a candidate server. The gateway row, with its tools, credential, owner, and traffic, becomes the inventory record.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Shadow MCP detection fills the unsanctioned half.&lt;/strong&gt; &lt;a href="https://www.speakeasy.com/blog/release-shadow-mcp-detection" rel="noopener noreferrer"&gt;Shadow MCP detection&lt;/a&gt; recognizes, from instrumented agent traffic, when a session connects to a server outside the sanctioned setup, and surfaces it by URL, host, or resolved identity. Allow and deny rules then apply at runtime. If policy evaluation is unavailable, the organization's fail-open or fail-closed setting determines whether the call proceeds; explicit denials and invalid credentials still block.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Identity and credentials sit on the gateway.&lt;/strong&gt; Upstream auth is provisioned per server through one-click OAuth on the user's own identity, a managed credential stored once, or header pass-through. Each inventory row carries a named owner and a known auth method.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The MCP inventory is built from live traffic, so it stays useful between reviews. To see the inventory the platform maintains for the servers your agents already call, &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is an MCP inventory?
&lt;/h3&gt;

&lt;p&gt;An MCP inventory is a continuously updated record of every Model Context Protocol server in an organization: the tools each server exposes, the credentials it authenticates with, who owns it, which agents may call it, and whether it is sanctioned. It records servers and tools, while an agent inventory records agents. Allowlists, tool-level least privilege, and audit evidence depend on it.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is the difference between an MCP inventory and an agent inventory?
&lt;/h3&gt;

&lt;p&gt;An MCP inventory records the MCP servers in use, the tools each one exposes, and the credentials behind it. An agent inventory records the agents themselves, the identity each runs as, and who owns it. The registers answer different questions and link to each other: a server record lists the agents allowed to call it, and an agent record lists the servers it connects to.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is the public MCP Registry an MCP inventory?
&lt;/h3&gt;

&lt;p&gt;The official MCP Registry at registry.modelcontextprotocol.io is a catalog of servers anyone can publish. An enterprise MCP inventory records the servers actually running in one organization: local stdio servers in dotfiles, remote vendor URLs, internal generated servers, and hosted connectors. The registry supplies candidates for the sanctioned list, while the inventory records which of them agents call.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does an MCP inventory relate to shadow MCP?
&lt;/h3&gt;

&lt;p&gt;Shadow MCP means unsanctioned servers operating outside formal security governance. Discovery surfaces those servers one finding at a time. The MCP inventory turns each finding into a record with an owner and a disposition, alongside every sanctioned server. Together, those records give the organization a tool-layer view.&lt;/p&gt;

&lt;h3&gt;
  
  
  What fields should an MCP inventory record capture?
&lt;/h3&gt;

&lt;p&gt;Each record should capture the server's identity (a URL for a remote server, a package specifier for a local one), its transport (stdio or remote), the tools it exposes with their read-only or destructive disposition, the credential it authenticates with and whose identity that credential carries, a human owner of record, the agents and clients allowed to connect, its sanctioned status, and when it was last seen serving a call.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does an MCP inventory stay current?
&lt;/h3&gt;

&lt;p&gt;A spreadsheet compiled for an audit describes last quarter by the time anyone reads it, because wiring up a new server is a config-file edit. A live inventory uses two feeds: hooks inside coding agents observe every server a session connects to, including unsanctioned ones, and a gateway on the sanctioned path records server and tool-call traffic. Together, they keep the last-seen field current.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What is an MCP registry?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Mon, 16 Nov 2026 08:31:16 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-an-mcp-registry-3ida</link>
      <guid>https://dev.to/speakeasy/what-is-an-mcp-registry-3ida</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/mcp-registry" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  MCP registry
&lt;/h2&gt;

&lt;p&gt;An MCP registry is a catalog of Model Context Protocol (MCP) server metadata.&lt;br&gt;
It names the server, its install location, and its connection method. The&lt;br&gt;
official registry lists publicly accessible servers. It does not record which&lt;br&gt;
servers your organization actually runs or accept private ones.&lt;/p&gt;

&lt;p&gt;An MCP registry supplies discovery metadata. It records what exists and, in its private form, what is approved to exist. It does not show what runs on employee laptops, which credentials those servers hold, or which agents call them. Those records belong to an &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;MCP inventory&lt;/a&gt;, while enforcement belongs to an &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt;. A registry, inventory, and gateway answer different security questions.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is an MCP registry?
&lt;/h2&gt;

&lt;p&gt;An MCP registry, also called an MCP server registry, stores metadata about MCP servers, and clients, marketplaces, and platform teams query it to discover them. The &lt;a href="https://modelcontextprotocol.io/registry/about" rel="noopener noreferrer"&gt;official registry documentation&lt;/a&gt; describes each entry as a standardized &lt;code&gt;server.json&lt;/code&gt; record containing the server's unique name in reverse-DNS format, such as &lt;code&gt;io.github.user/server-name&lt;/code&gt;, a pointer to where the server lives, such as an npm package name or a remote server URL, execution instructions such as command-line arguments and environment variables, and discovery data such as a description and capabilities.&lt;/p&gt;

&lt;p&gt;The registry hosts none of the code. Packages can stay on npm, PyPI, or Docker Hub, while remote servers can remain on the vendor's infrastructure. The registry entry points at them. Consumers read the catalog through a &lt;a href="https://modelcontextprotocol.io/registry/registry-aggregators" rel="noopener noreferrer"&gt;REST API&lt;/a&gt;, with &lt;code&gt;GET /v0.1/servers&lt;/code&gt; at &lt;code&gt;registry.modelcontextprotocol.io&lt;/code&gt; returning the full list for downstream systems to ingest.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is there an official MCP registry?
&lt;/h2&gt;

&lt;p&gt;Yes. The official MCP Registry at &lt;a href="https://registry.modelcontextprotocol.io/" rel="noopener noreferrer"&gt;registry.modelcontextprotocol.io&lt;/a&gt; launched in &lt;a href="https://blog.modelcontextprotocol.io/posts/2025-09-08-mcp-registry-preview/" rel="noopener noreferrer"&gt;preview in September 2025&lt;/a&gt; as part of the MCP open-source project, backed by Anthropic, GitHub, PulseMCP, and Microsoft. It remains in preview as of its &lt;a href="https://modelcontextprotocol.io/registry/about" rel="noopener noreferrer"&gt;about page&lt;/a&gt;, which warns that breaking changes or data resets may occur before general availability, so treat any dependency on it as a dependency on a preview service.&lt;/p&gt;

&lt;p&gt;Its scope is narrower than the name suggests. The boundaries matter for anyone evaluating it as a governance control.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Namespace authentication does not verify server code.&lt;/strong&gt; Publishing under &lt;code&gt;io.github.*&lt;/code&gt; requires the matching GitHub account, and domain namespaces require DNS or HTTP verification. The registry performs no security scanning of server code. It &lt;a href="https://modelcontextprotocol.io/registry/about" rel="noopener noreferrer"&gt;delegates scanning&lt;/a&gt; to package registries and downstream aggregators, and its own moderation is limited to takedowns of spam and malicious entries.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Public servers only.&lt;/strong&gt; The registry &lt;a href="https://modelcontextprotocol.io/registry/about" rel="noopener noreferrer"&gt;does not support private servers&lt;/a&gt;, meaning anything on an internal hostname or a private package registry. The official recommendation for those is to host your own private MCP registry.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not built for self-hosting.&lt;/strong&gt; The official codebase is not designed to be run internally, and the maintainers state they cannot support forks.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Aggregators are the intended readers.&lt;/strong&gt; The registry expects marketplaces and subregistries to consume its API and serve host applications, rather than host applications querying it directly, with the &lt;a href="https://modelcontextprotocol.io/registry/faq" rel="noopener noreferrer"&gt;OpenAPI spec&lt;/a&gt; as the shared interface.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What are the third-party MCP registries?
&lt;/h2&gt;

&lt;p&gt;Third-party MCP registries have multiplied since the official preview launched. The official project's &lt;a href="https://modelcontextprotocol.io/registry/faq" rel="noopener noreferrer"&gt;naming convention&lt;/a&gt; defines an "MCP registry" in lowercase as any third-party service implementing the MCP Registry API. For a CISO, sorting them into three kinds is more useful than enumerating them because each kind answers a different procurement question.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Registry&lt;/th&gt;
&lt;th&gt;Kind&lt;/th&gt;
&lt;th&gt;What it is&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;a href="https://github.blog/ai-and-ml/github-copilot/meet-the-github-mcp-registry-the-fastest-way-to-discover-mcp-servers/" rel="noopener noreferrer"&gt;GitHub MCP Registry&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Aggregator&lt;/td&gt;
&lt;td&gt;Curated directory launched September 2025 with one-click install into VS Code. GitHub is working with Anthropic and the MCP Steering Committee so servers published to the official registry flow into it.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;a href="https://www.pulsemcp.com/" rel="noopener noreferrer"&gt;PulseMCP&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Aggregator&lt;/td&gt;
&lt;td&gt;Community server directory whose maintainer helped build the official registry and serves as one of its maintainers.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Azure API Center&lt;/td&gt;
&lt;td&gt;Cloud catalog&lt;/td&gt;
&lt;td&gt;Registers remote and local MCP servers as inventory assets and exposes them at a registry endpoint in the official &lt;code&gt;/v0.1/servers&lt;/code&gt; format.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;a href="https://jfrog.com/ai-catalog/mcp-registry/" rel="noopener noreferrer"&gt;JFrog AI Catalog MCP Registry&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Cloud catalog&lt;/td&gt;
&lt;td&gt;Manages MCP servers as artifacts inside the JFrog platform, with policy-based blocking of unvetted servers and role-based access to tools.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Kong MCP Registry&lt;/td&gt;
&lt;td&gt;Enterprise product&lt;/td&gt;
&lt;td&gt;Registry entity inside Kong Konnect, in Tech Preview, paired with the Kong AI Gateway for runtime enforcement.&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;a href="https://portkey.ai/docs/product/mcp-gateway/mcp-registry" rel="noopener noreferrer"&gt;Portkey MCP Registry&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;Enterprise product&lt;/td&gt;
&lt;td&gt;Central catalog of MCP server integrations inside the Portkey MCP gateway, with per-workspace access control and per-server authentication.&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The evaluation question is which kind you are looking at. An aggregator improves discovery of public servers. A cloud catalog folds MCP servers into an inventory you already keep. An enterprise product adds governance to the same OpenAPI shape. Whichever kind you pick, the record of what is actually running on your endpoints still comes from an &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;MCP inventory&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is a private or central MCP registry?
&lt;/h2&gt;

&lt;p&gt;A private MCP registry is the catalog an organization hosts for the servers the public one will not take. The official documentation &lt;a href="https://modelcontextprotocol.io/registry/about" rel="noopener noreferrer"&gt;recommends&lt;/a&gt; that servers on internal hostnames or private package registries be published to a private MCP registry the organization runs itself, implementing the same OpenAPI spec so existing clients and host applications can consume it without custom work.&lt;/p&gt;

&lt;p&gt;A central MCP registry is the organization-wide version of the same idea, and it is the phrase that matters in a security review. It identifies which servers employees may install, who owns each one, and which environments see them, instead of relying on per-developer &lt;code&gt;mcp.json&lt;/code&gt; files. The official preview announcement &lt;a href="https://blog.modelcontextprotocol.io/posts/2025-09-08-mcp-registry-preview/" rel="noopener noreferrer"&gt;describes this pattern&lt;/a&gt; for private subregistries inside enterprises that use the official registry as their upstream data source.&lt;/p&gt;

&lt;h3&gt;
  
  
  What a central MCP registry records
&lt;/h3&gt;

&lt;p&gt;A central registry records that a server is approved and who owns the approval, not which employees actually run it or whether an unapproved server has appeared. Filling that gap takes a continuously updated record of what's running, which is the job of an &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;MCP inventory&lt;/a&gt;; the unsanctioned population it surfaces is &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How is an MCP registry different from an MCP gateway?
&lt;/h2&gt;

&lt;p&gt;Once a client has an entry's connection details, the registry's job is done. Everything that happens on the tool-call path, identity checks, allowlists, audit logging, data controls, belongs to an &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt;, the layer where &lt;a href="https://www.speakeasy.com/resources/mcp-security" rel="noopener noreferrer"&gt;MCP security&lt;/a&gt; controls actually execute.&lt;/p&gt;

&lt;h3&gt;
  
  
  When should an organization use an MCP registry and gateway together?
&lt;/h3&gt;

&lt;p&gt;Each works without the other, with predictable gaps:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A registry alone cannot block a call. An approved-server catalog without an enforcement point is a policy document.&lt;/li&gt;
&lt;li&gt;A gateway alone governs traffic when servers are added by URL. It does not provide discovery of public servers or a sanctioned catalog for employees to browse.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Together, the registry defines what is approved and the gateway makes that approval binding. That pairing explains why the enterprise products in the table include both.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/mcp-registry#when-should-an-organization-use-an-mcp-registry-and-gateway-together" rel="noopener noreferrer"&gt;View the diagram separating discovery, runtime enforcement, and observation on Speakeasy&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  How does Speakeasy use the official MCP Registry?
&lt;/h2&gt;

&lt;p&gt;The Speakeasy AI Control Plane uses the official MCP Registry as the source for its &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/catalog" rel="noopener noreferrer"&gt;catalog&lt;/a&gt;. The Catalog page, under Connect &amp;gt; Catalog in the dashboard, lets teams browse, search, and filter registry entries, then install the servers they approve into a project. Viewing requires the &lt;code&gt;project:read&lt;/code&gt; or &lt;code&gt;mcp:write&lt;/code&gt; scope and adding a server requires &lt;code&gt;mcp:write&lt;/code&gt;, so a Member can browse the catalog while installs stay with holders of the write scope, which the default Admin role includes.&lt;/p&gt;

&lt;p&gt;An install creates a &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/remote-servers" rel="noopener noreferrer"&gt;remote MCP server&lt;/a&gt; on the platform. The vendor keeps running the server. The platform registers its URL and proxies sessions to it, adding authentication, per-tool access control, and tool logs without copying or re-hosting the server. One catalog entry can create several servers when it publishes multiple remote endpoints. Only streamable HTTP endpoints are accepted, and entries marked Manual Setup require configuration outside the one-click path.&lt;/p&gt;

&lt;p&gt;The registry feed covers the sanctioned side only. Servers employees wire up on their own are surfaced by &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP detection&lt;/a&gt;, which works from agent hooks and live traffic rather than from any registry, and both populations land in the &lt;a href="https://www.speakeasy.com/resources/mcp-inventory" rel="noopener noreferrer"&gt;MCP inventory&lt;/a&gt; the platform maintains. The official registry supplies the catalog of what exists, the platform turns the approved subset into a governed and auditable path, and the gateway and inventory cover what is actually running. To see the catalog and the inventory against your own estate, visit the &lt;a href="https://www.speakeasy.com/product/ai-control-plane" rel="noopener noreferrer"&gt;AI Control Plane&lt;/a&gt; or &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Organization administrators can also search the catalog, register servers, and check readiness from an agent through the &lt;a href="https://www.speakeasy.com/blog/release-platform-mcp" rel="noopener noreferrer"&gt;Speakeasy Platform MCP&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is an MCP registry?
&lt;/h3&gt;

&lt;p&gt;An MCP registry is a catalog of Model Context Protocol (MCP) server metadata. It names the server, its install location, and its connection method. The official registry lists publicly accessible servers. It does not record which servers your organization actually runs or accept private ones. Each entry points at the package registry or remote endpoint where the server lives. A registry answers what exists and how to install it. An inventory records what is currently running.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is there an official MCP registry?
&lt;/h3&gt;

&lt;p&gt;Yes. The official MCP Registry at registry.modelcontextprotocol.io launched in preview in September 2025 and is backed by Anthropic, GitHub, PulseMCP, and Microsoft. It is still in preview as of its documentation, which warns that breaking changes or data resets may occur before general availability. It authenticates namespaces, hosts metadata in a standardized server.json format, and delegates security scanning of server code to package registries and downstream aggregators.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are the third-party MCP registries?
&lt;/h3&gt;

&lt;p&gt;Third-party MCP registries group into three kinds. Aggregators and marketplaces such as the GitHub MCP Registry and PulseMCP consume the official registry's API and add curation, ratings, and one-click installs. Cloud and vendor catalogs such as Azure API Center and the JFrog AI Catalog MCP Registry inventory MCP servers inside a platform the organization already runs. Enterprise registry products such as the Kong MCP Registry, in Tech Preview, and the Portkey MCP Registry implement the same OpenAPI shape and add access control and governance on top.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is a private or central MCP registry?
&lt;/h3&gt;

&lt;p&gt;A private MCP registry is a catalog one organization hosts for servers the official registry does not accept. Internal hostnames and private package registries stay out of the public catalog by design. The official documentation recommends hosting a registry that implements the same OpenAPI spec so clients can consume it. A central MCP registry is the organization-wide version of the same idea. It records which servers employees may install, who owns each one, and which environments see them instead of leaving the information in individual mcp.json files. Both are catalogs of metadata rather than live inventories of what is running.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is an MCP registry different from an MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A registry handles discovery. It records server metadata, install targets, and namespaces outside the tool-call path. An MCP gateway handles enforcement. It terminates tool calls and applies identity, allowlists, audit logging, and data controls on every request. A registry cannot block a call, and a gateway works without a registry when servers are added by URL. Organizations that need both a sanctioned catalog and runtime enforcement can run the two together, with the registry feeding the gateway's allowlist.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does Speakeasy use the official MCP Registry?
&lt;/h3&gt;

&lt;p&gt;The Speakeasy AI Control Plane uses the official MCP Registry to power its catalog. Teams browse, search, and filter registry entries from Connect &amp;gt; Catalog in the dashboard, then install the servers they approve. Each install registers the vendor's remote server URL and proxies sessions to it. The platform adds authentication, per-tool access control, and tool logs without copying or re-hosting the server. A separate detection feed built on agent hooks and live traffic surfaces unsanctioned servers.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What is an MCP tunnel?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Fri, 13 Nov 2026 13:11:37 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-an-mcp-tunnel-33gn</link>
      <guid>https://dev.to/speakeasy/what-is-an-mcp-tunnel-33gn</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/mcp-tunnels" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  MCP tunnel
&lt;/h2&gt;

&lt;p&gt;An MCP tunnel is an outbound-only connection that lets an AI agent or platform&lt;br&gt;
outside a private network reach a specific Model Context Protocol (MCP) server&lt;br&gt;
inside it. A small tunnel client runs next to the server and dials out over TLS&lt;br&gt;
to the endpoint that needs to reach it. Tool calls travel back down that&lt;br&gt;
connection, and nothing inside the network accepts inbound traffic.&lt;/p&gt;

&lt;p&gt;The MCP servers most worth connecting to agents are usually the ones an organization can least afford to expose. Because they remain private, a Postgres MCP tool that queries live data, an admin tool that can edit a user's plan, or a wrapper for a company API stays hidden. It has no public URL, and its firewall exists to keep it that way.&lt;/p&gt;

&lt;p&gt;They run differently. &lt;a href="https://www.speakeasy.com/resources/background-agents" rel="noopener noreferrer"&gt;Background agents&lt;/a&gt; operate using vendors' cloud infrastructure or separate VMs, with no access to company systems, while MCP apps like Claude and ChatGPT reach services through URLs. A tool running outside the network cannot access servers that exist only within it.&lt;/p&gt;

&lt;p&gt;Before tunnels, fixing that split came down to one of 2 things. The machine gained an open endpoint, the risk the firewall was meant to block, or the software got in through a VPN or peering setup, touching much more than one machine. With MCP, a link lets the app reach only that server and no other, while the server remains hidden.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does an MCP tunnel work?
&lt;/h2&gt;

&lt;p&gt;A tunnel is a reverse connection. A tunnel client runs on a machine already connected to the MCP server and dials outbound, rather than accepting inbound connections:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The client opens one outbound connection.&lt;/strong&gt; It connects over HTTPS or a secure WebSocket to a tunnel endpoint outside the network. Most corporate firewalls already allow outbound traffic on port 443, so a tunnel typically requires no firewall changes and no network team coordination.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tool calls travel through that connection.&lt;/strong&gt; When an agent calls a tool, the request travels down the connection the client opened, the client forwards it locally to the MCP server, and the response returns the same way. Concurrent requests are multiplexed as separate streams over the single connection.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The server remains hidden inside.&lt;/strong&gt; There is no inbound firewall rule, no public hostname, and no port forward. The only path to the server is the connection the client chose to open, and stopping the client closes it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Two details separate a well-built tunnel from an ad hoc one. It should only send data to the nearby IP given when configured on startup; then its endpoint, once it dials, cannot redirect it toward another machine in the LAN. It must authenticate against the endpoint through a secret the team may rotate, so an exposed secret becomes invalid with no host changes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where do MCP tunnels connect?
&lt;/h2&gt;

&lt;p&gt;A tunnel connects one private server to one endpoint. That endpoint is the thing trying to contact the server:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;An agent vendor's platform.&lt;/strong&gt; &lt;a href="https://claude.com/blog/claude-managed-agents-updates" rel="noopener noreferrer"&gt;Anthropic&lt;/a&gt; and &lt;a href="https://developers.openai.com/blog/connect-private-mcp-servers-to-openai-products" rel="noopener noreferrer"&gt;OpenAI&lt;/a&gt; each ship an MCP tunnel so their hosted products can reach private servers. A Claude tunnel serves Managed Agents and the Messages API, and an OpenAI tunnel serves ChatGPT, Codex, and other OpenAI products.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;An MCP gateway.&lt;/strong&gt; The tunnel terminates at an &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt;, which then serves the private server to every agent surface behind it and applies its identity, permission, and logging controls to tunneled traffic.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A developer's own tooling.&lt;/strong&gt; A developer building an MCP server on a laptop can tunnel it so a hosted client can call it before it is deployed anywhere.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The decision counts because a link provides connectivity, without control. A tunneled host follows rules set at the endpoint it reaches. Setting up one path per vendor means every vendor gets a separate endpoint on the system, so none can see what another dialed. Routing through a gateway means the firm gets a single user on each host, plus one spot logging every request.&lt;/p&gt;

&lt;h2&gt;
  
  
  When does an organization need an MCP tunnel?
&lt;/h2&gt;

&lt;p&gt;Software on a device within a network can connect directly to a private host. A tunnel becomes necessary when an agent and a host operate across a network boundary. Common scenarios include:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Hosted agents needing internal systems.&lt;/strong&gt; A background agent on vendor infrastructure requires access to an issue tracker behind SSO, a staging database, or an internal API that only exist inside the network.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Internal APIs behind the firewall.&lt;/strong&gt; APIs built for callers on the corporate network cannot be exposed publicly without inverting the security model they were designed for.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Development servers on localhost.&lt;/strong&gt; An MCP server under development has no address outside the developer's machine. A tunnel is the only way for a hosted agent to call it before deployment.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;VPC-only services.&lt;/strong&gt; Databases and internal admin platforms are intentionally private, and a tunnel provides access without changing that privacy posture.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;On-premises systems in restricted networks.&lt;/strong&gt; Where policy prohibits inbound traffic, an outbound-only connection is the only pattern that satisfies security review.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What value do MCP tunnels provide?
&lt;/h2&gt;

&lt;p&gt;An MCP tunnel offers distinct advantages over other connectivity approaches:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Private servers become usable by agents.&lt;/strong&gt; The most valuable servers, the ones holding sensitive data or controlling consequential actions, are private for a reason. A tunnel makes them available to agents without forcing organizations to decide whether they are safe to publish.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No new inbound firewall exposure.&lt;/strong&gt; No inbound rule or port forward is added, and the private server needs no public hostname. The tunnel still creates a new access path, so the endpoint it connects to needs appropriate authentication and policy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The network path is scoped to one server.&lt;/strong&gt; The client forwards to only one configured local address; it does not put the agent on the network. Access to that server still depends on the controls where the tunnel terminates.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deployment is lightweight.&lt;/strong&gt; The tunnel client is a single process or container running anywhere the server is reachable. Since outbound port 443 is almost universally open, most deployments require no firewall changes or network team coordination.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Control remains with the organization.&lt;/strong&gt; The tunnel exists only while the client runs. The authentication key can be rotated, and the local destination address is fixed at launch time. Nothing external can widen the tunnel's scope.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Governance placement is a choice.&lt;/strong&gt; Terminate the tunnel at a policy gateway and the private server gains identity, permissions, and an audit trail without leaving the network. Terminate it at an agent directly and the server becomes reachable but ungoverned, which creates a &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP&lt;/a&gt; problem.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  MCP tunnel vs VPN: Key differences
&lt;/h2&gt;

&lt;p&gt;An MCP tunnel and a VPN both bridge networks, but serve different purposes and operate at different scopes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;VPN scope:&lt;/strong&gt; Gives a client network-layer access to the services permitted by its routes and access rules&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;MCP tunnel scope:&lt;/strong&gt; Connects a single MCP server to a single endpoint; the tunnel client forwards only to one configured local address&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;VPN exposure:&lt;/strong&gt; Depends on routes, access rules, and device policy; a VPN can be tightly scoped&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;MCP tunnel exposure:&lt;/strong&gt; The tunnel forwards only to its configured MCP endpoint; the terminating platform decides who may call it&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;VPN setup:&lt;/strong&gt; Requires VPN client software, credentials, and ongoing session management&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;MCP tunnel setup:&lt;/strong&gt; Single outbound connection on port 443, usually requires no firewall changes&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For organizations needing agents to reach a specific private MCP server without giving them a VPN session, a tunnel narrows the application path. It does not replace authentication and authorization at the endpoint.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does Speakeasy provide MCP tunnels?
&lt;/h2&gt;

&lt;p&gt;In the &lt;a href="https://www.speakeasy.com/resources/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt;, a tunneled server operates as a source type. A lightweight proxy runs alongside your internal host, establishing one outbound WebSocket connection to the platform that multiplexes per-request streams through it. It forwards only to the local MCP endpoint pinned at launch, reconnects automatically with backoff if the link fails, and authenticates with a token displayed once, retained only as a hash, and rotated to revoke all active connections.&lt;/p&gt;

&lt;p&gt;The tunnel creates a persistent bridge between your private server and the external platform, with requests and responses flowing through the outbound connection established by your network, as shown in the linked diagram:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/mcp-tunnels#how-does-speakeasy-provide-mcp-tunnels" rel="noopener noreferrer"&gt;View the MCP tunnel architecture diagram on Speakeasy&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;After setup, each tunneled host behaves the same as any other entry in that registry. It receives a cloud URL, runs via the same proxy layer as outside machines, and gains authentication, per-role app permissions, metering, and group entry. Calls appear in &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/observe/tool-logs" rel="noopener noreferrer"&gt;tool logs&lt;/a&gt; with the Tunneled MCP target type next to everything else the organization runs. One connection leaves it reachable to Claude, ChatGPT, the Responses API, or an organization's own tools via that governed route, not one per vendor.&lt;/p&gt;

&lt;p&gt;It appears as an option inside the dashboard. Create one tunneled MCP server from &lt;strong&gt;Sources &amp;gt; Connect&lt;/strong&gt;; use its Docker, CLI, or Kubernetes setup where it can access that server, and it appears in the registry once it dials outward. The &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/tunneled-servers" rel="noopener noreferrer"&gt;tunneled MCP servers&lt;/a&gt; docs cover configuration and monitoring, and the &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/guides/deploy-tunnel-agent-aws" rel="noopener noreferrer"&gt;AWS deployment guide&lt;/a&gt; walks through EKS, ECS Fargate, and EC2.&lt;/p&gt;

&lt;p&gt;To reach isolated hosts from your apps without exposing those, &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk with us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is an MCP tunnel?
&lt;/h3&gt;

&lt;p&gt;An MCP tunnel is an outbound-only connection that lets an AI agent or platform outside a private network reach a specific Model Context Protocol (MCP) server inside it. A small tunnel client runs next to the server, dials out over TLS to the endpoint that needs to reach the server, and tool calls travel back down that connection. The server needs no inbound firewall rule, no public hostname, and no port forward.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why do organizations need MCP tunnels?
&lt;/h3&gt;

&lt;p&gt;The MCP servers most worth connecting to agents, such as a database server that queries production or an admin server for an internal system, run inside private networks. Agents increasingly run outside those networks, on a vendor's hosted infrastructure or an isolated VM. A tunnel lets the agent reach the server without the organization exposing it publicly, opening a firewall port, or standing up a VPN for the agent.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does an MCP tunnel work?
&lt;/h3&gt;

&lt;p&gt;It is a reverse tunnel. A tunnel client runs on a host that already reaches the MCP server and opens one outbound connection, usually over HTTPS or a secure WebSocket on port 443, to the endpoint outside the network. When an agent calls a tool, the request is sent down that established connection, the client forwards it to the server locally, and the response returns the same way. Concurrent requests are multiplexed as separate streams over the one connection.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is an MCP tunnel the same as a VPN?
&lt;/h3&gt;

&lt;p&gt;No. A VPN joins a machine to a network, so a connected agent could reach anything on it. An MCP tunnel connects one MCP server to one endpoint, and the tunnel client only forwards traffic to the local address it was configured with. The exposure is a single server rather than a network segment, and stopping the client closes the only path in.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is an MCP tunnel the same as an MCP gateway?
&lt;/h3&gt;

&lt;p&gt;No. A tunnel is connectivity. It carries MCP traffic to a server that has no public address. An MCP gateway is a policy point that applies identity, permissions, and logging to tool calls. A tunnel can terminate at a gateway, at an agent vendor's platform, or at a developer's own tooling, and the governance a tunneled server gets depends on where the tunnel ends, not on the tunnel itself.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does Speakeasy provide MCP tunnels?
&lt;/h3&gt;

&lt;p&gt;The Speakeasy AI Control Plane connects a private server through a tunneled MCP source. A lightweight tunnel agent runs next to the server and opens a single outbound WebSocket connection to the platform, which multiplexes per-request streams back down that connection. The server then behaves like any other server in the registry, with a hosted URL, authentication, per-role tool permissions, usage metering, and tool logs. One tunnel makes the server reachable from Claude, ChatGPT, the Responses API, or an organization's own agents.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What is MCP gateway bypass?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Fri, 13 Nov 2026 12:42:03 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-mcp-gateway-bypass-33l9</link>
      <guid>https://dev.to/speakeasy/what-is-mcp-gateway-bypass-33l9</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/mcp-gateway-bypass" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  MCP gateway bypass
&lt;/h2&gt;

&lt;p&gt;MCP gateway bypass is any route by which an AI agent reaches a Model Context&lt;br&gt;
Protocol (MCP) server, or the API behind it, without passing through the&lt;br&gt;
organization's MCP gateway. The gateway's authentication, tool policy, and&lt;br&gt;
audit log apply only to requests it receives, so a call that takes another&lt;br&gt;
route carries none of them.&lt;/p&gt;

&lt;p&gt;An &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt; is the policy point between agents and tools. It authenticates the caller, decides which tools they may use, and logs every call. All of that assumes the call arrives. Connecting an agent to a server takes one line of client configuration, and Cloudflare describes how it usually happens: an employee finds a server in a repository, a product guide, or a message from a colleague and adds it to their MCP client directly. Nothing about that connection passes the gateway, and the gateway has no record that it exists.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A gateway governs the traffic it receives, so preventing MCP gateway bypass means making every other route fail, and each route is closed by a different layer.&lt;/strong&gt; Two of those layers live in the gateway. The others live on the network, on the device, and at the upstream provider, and one route stays open whatever you deploy: a device the organization doesn't manage.&lt;/p&gt;
&lt;h2&gt;
  
  
  The four bypass paths and the layer that closes each
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Local stdio servers.&lt;/strong&gt; The client launches the server as a subprocess, and no MCP traffic leaves the machine. Closed by client hooks and device-level discovery.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Direct remote URLs.&lt;/strong&gt; The client connects to a remote server, approved or not, at its own address instead of through the gateway. Closed by network egress policy and by an origin that rejects requests that didn't come through the gateway.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Personal tokens.&lt;/strong&gt; The user holds a working credential for the upstream service and uses it from any client or script. Closed by credential custody at the gateway and by the provider's own token policy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Unmanaged clients.&lt;/strong&gt; An MCP client nobody reviewed, or a reviewed client that lost the organization's configuration, connects to the organization's tools. Closed by client admission at the gateway and by device enrollment.&lt;/li&gt;
&lt;/ol&gt;
&lt;h2&gt;
  
  
  How is MCP gateway bypass different from shadow MCP?
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;Shadow MCP&lt;/a&gt; is a connection to an MCP server the organization has not approved. The OWASP MCP Top 10 catalogs it as &lt;a href="https://owasp.org/www-project-mcp-top-10/" rel="noopener noreferrer"&gt;MCP09:2025 Shadow MCP Servers&lt;/a&gt;. Gateway bypass is the wider set. It includes every shadow MCP connection, and it also includes connections to approved servers that skip the gateway placed in front of them.&lt;/p&gt;

&lt;p&gt;Cloudflare makes the same split in its August 2026 post on MCP detection. It calls the second case portal bypass: an approved server sits behind an MCP Portal, and an employee connects to its upstream URL anyway, skipping the Portal's access policy, tool catalog, data loss prevention, and audit trail. Cloudflare's conclusion is that shadow MCP needs a network control, while portal bypass needs that network control plus an origin that can reject direct requests.&lt;/p&gt;

&lt;p&gt;Approving a server therefore doesn't close the bypass. An organization can put every approved server behind a gateway and still find calls to those servers going direct, because the old client configuration still works. &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;Shadow AI&lt;/a&gt; is the broader category again, covering models, agents, skills, and MCP servers used outside the security perimeter. This article covers only the routes around the MCP gateway.&lt;/p&gt;
&lt;h2&gt;
  
  
  Why can't the gateway stop bypass by itself?
&lt;/h2&gt;

&lt;p&gt;The gateway is a server, so it can refuse a request it receives but has no view of a request sent somewhere else. Three of the four bypass paths never produce a request to the gateway.&lt;/p&gt;

&lt;p&gt;The local stdio path is the clearest case. The &lt;a href="https://modelcontextprotocol.io/specification/2026-07-28/basic/transports" rel="noopener noreferrer"&gt;MCP transports specification&lt;/a&gt; defines stdio as a transport where the client launches the server as a subprocess and exchanges messages over standard input and output. A server configured that way runs entirely on the developer's machine. If it calls an API, that call is ordinary HTTPS from the laptop to the API, with nothing in it that says MCP.&lt;/p&gt;

&lt;p&gt;Direct remote connections and personal tokens are the same problem one step out. The request goes from the client to a server or API the gateway doesn't front, so the gateway's policy never runs. Unmanaged clients are the exception, because they do connect to the gateway. That makes them the one path a gateway can close at the point of connection.&lt;/p&gt;

&lt;p&gt;Deploying a gateway gives approved use a governed path. The other layers make the ungoverned paths fail, or at least show up somewhere a security team can see them.&lt;/p&gt;
&lt;h2&gt;
  
  
  How do agents route around an MCP gateway?
&lt;/h2&gt;

&lt;p&gt;Each path looks different in practice and leaves different evidence:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Bypass path&lt;/th&gt;
&lt;th&gt;What it looks like&lt;/th&gt;
&lt;th&gt;What the gateway sees&lt;/th&gt;
&lt;th&gt;Layer that closes it&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Local stdio server&lt;/td&gt;
&lt;td&gt;A client config entry that runs a command such as &lt;code&gt;npx some-mcp-server&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;Nothing&lt;/td&gt;
&lt;td&gt;Client hooks and device-level discovery&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Direct remote URL&lt;/td&gt;
&lt;td&gt;A server URL pasted into a client, either an unapproved server or the upstream address of an approved one&lt;/td&gt;
&lt;td&gt;Nothing&lt;/td&gt;
&lt;td&gt;Network egress policy, plus an origin that rejects direct requests&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Personal token&lt;/td&gt;
&lt;td&gt;A personal access token or API key in a client config or environment variable&lt;/td&gt;
&lt;td&gt;Nothing&lt;/td&gt;
&lt;td&gt;Credential custody at the gateway, plus provider token policy&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Unmanaged client&lt;/td&gt;
&lt;td&gt;An unreviewed MCP client, or a client switched to a personal account that dropped the organization's plugins and hooks&lt;/td&gt;
&lt;td&gt;An authorization from an unfamiliar client, which it may admit&lt;/td&gt;
&lt;td&gt;Client admission at the gateway, plus device enrollment&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/mcp-gateway-bypass#:~:text=An%20authorization%20from%20an%20unfamiliar%20client%2C%20which%20it%20may%20admit" rel="noopener noreferrer"&gt;View the MCP gateway bypass paths diagram on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The personal-token path outlives the others. A user who connected a server directly last quarter may still hold the token. Removing the server from their client config doesn't revoke it, and the &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP guide&lt;/a&gt; makes the same point about closing findings: a config entry removed without addressing the credential leaves the connection easy to restore.&lt;/p&gt;

&lt;p&gt;The unmanaged-client path has a quieter version that falls under configuration drift, which Speakeasy's &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/reference/device-agent" rel="noopener noreferrer"&gt;device agent documentation&lt;/a&gt; calls the most common bypass on developer laptops. A user switches their AI agent to a personal account and silently loses the organization's plugins, telemetry, and guardrails. The client is the same one the organization approved, but its configuration is not.&lt;/p&gt;
&lt;h2&gt;
  
  
  How do you prevent MCP gateway bypass?
&lt;/h2&gt;

&lt;p&gt;Five control layers cover the four paths: network egress, the origin server, credential custody, client admission, and device-level discovery. None covers every path, and each has a documented blind spot.&lt;/p&gt;
&lt;h3&gt;
  
  
  Network egress: block direct MCP traffic on managed paths
&lt;/h3&gt;

&lt;p&gt;A secure web gateway sees HTTP traffic after it leaves the client, which makes it the widest view of remote MCP connections from managed devices. Cloudflare Gateway now classifies TLS-inspected requests as MCP when they carry the &lt;code&gt;MCP-Protocol-Version&lt;/code&gt; header, and exposes that as a beta Is MCP selector in HTTP policies. Combined with a Traffic Source selector that marks requests proxied by an MCP Portal, one rule blocks MCP traffic that didn't come through the Portal:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;experimental.is_mcp == true and not traffic.onramp in ("mcp_portal")
Action: Block
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Docker documents the same two-layer pattern for its sandboxes. Its MCP access policies govern only server registration and requests handled by Docker's MCP gateway. A direct connection from inside a sandbox to a remote MCP server is outbound sandbox traffic, so Docker's guidance is to block the server in MCP policy and block its network destination in network policy.&lt;/p&gt;

&lt;p&gt;Egress control has three limits, and Cloudflare states all of them. Detection requires TLS decryption. Local stdio servers, off-network connections, and traffic exempted from inspection stay outside the view. And legacy MCP clients send their first &lt;code&gt;initialize&lt;/code&gt; request without the protocol header, so the first request to a previously unknown endpoint may not be classified. An egress rule is therefore a control for managed devices on managed networks, and the rest of the stack has to assume it can be absent.&lt;/p&gt;

&lt;h3&gt;
  
  
  The origin: make direct requests fail at the server
&lt;/h3&gt;

&lt;p&gt;An egress rule stops the request on the way out. A locked-down origin stops it on arrival, which also covers devices the egress proxy never sees. Cloudflare's &lt;a href="https://developers.cloudflare.com/cloudflare-one/access-controls/ai-controls/mcp-portals/" rel="noopener noreferrer"&gt;MCP server portal documentation&lt;/a&gt; is direct about why this layer is needed: users blocked by a portal's Access policy can still connect to the server, and bypass that policy, by using its direct URL, unless Access is configured as the server's OAuth provider.&lt;/p&gt;

&lt;p&gt;For MCP servers the organization runs, the strongest origin control is to give the server no address anyone else can reach. An &lt;a href="https://www.speakeasy.com/resources/mcp-tunnels" rel="noopener noreferrer"&gt;MCP tunnel&lt;/a&gt; lets the gateway reach a server inside a private network over an outbound-only connection, with no inbound port open. The server can also check that a request came through the gateway. Speakeasy's private &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/tunneled-servers" rel="noopener noreferrer"&gt;tunneled servers&lt;/a&gt; receive a short-lived signed JWT in an &lt;code&gt;X-Speakeasy-Identity&lt;/code&gt; header on each forwarded request, which the server can verify and reject when absent.&lt;/p&gt;

&lt;p&gt;Private networking on the client side is a separate decision with its own costs. A &lt;a href="https://www.speakeasy.com/resources/private-mcp-gateway" rel="noopener noreferrer"&gt;private MCP gateway&lt;/a&gt; takes the gateway's own listener off the public internet, so only devices on the private network can reach it. That narrows who can reach the gateway, and it means hosted agents such as remote connectors in ChatGPT or Claude can't reach it at all. On Speakeasy, private network access runs over a Tailscale tailnet, requires an Enterprise organization, and is set per server as public only, public and private, or private only. The &lt;a href="https://www.speakeasy.com/resources/private-mcp-gateway" rel="noopener noreferrer"&gt;private MCP gateway guide&lt;/a&gt; covers when that trade is worth making. Speakeasy runs as a &lt;a href="https://www.speakeasy.com/resources/managed-vs-self-hosted-mcp-gateway" rel="noopener noreferrer"&gt;managed gateway&lt;/a&gt; by default, and self-hosting is available for Enterprise customers (&lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk to our team&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;For third-party SaaS servers, the origin belongs to the vendor, and the organization's options depend on what the vendor exposes.&lt;/p&gt;

&lt;h3&gt;
  
  
  Credential custody: keep the token in the gateway
&lt;/h3&gt;

&lt;p&gt;A direct connection to a SaaS MCP server works only if the user can authenticate to it. If the upstream credential lives in the gateway and never reaches the user, a direct connection has nothing to authenticate with. This is the cheapest layer to deploy, because it needs no software on any device.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices" rel="noopener noreferrer"&gt;MCP security best practices&lt;/a&gt; already forbid token passthrough, where a server accepts a client's token and forwards it upstream. Gateways implement the custody side in different ways. Kong's AI MCP OAuth2 plugin, part of AI Gateway Enterprise, validates that each token was issued for the target MCP server and doesn't forward client access tokens upstream by default. IBM's ContextForge stores per-user OAuth tokens encrypted, keyed by gateway and user, and keeps header passthrough off unless an operator enables it.&lt;/p&gt;

&lt;p&gt;Speakeasy's gateway holds &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/access/upstream-credentials" rel="noopener noreferrer"&gt;upstream credentials&lt;/a&gt; one of two ways: a token per user, obtained through the upstream's OAuth flow, encrypted at rest and refreshed by the platform, or &lt;a href="https://www.speakeasy.com/resources/shared-credentials-vs-user-delegated-identity" rel="noopener noreferrer"&gt;one shared credential&lt;/a&gt; per remote server, either a static key or an OAuth client credentials token. With &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/identity/identities/enterprise-managed-auth" rel="noopener noreferrer"&gt;Enterprise Managed Auth&lt;/a&gt; through Okta, the platform exchanges each person's Okta identity for the downstream token, so there's no per-service connect step and removing a person in Okta stops new access to every server that uses it.&lt;/p&gt;

&lt;p&gt;Custody stops new tokens from reaching users. It does nothing about tokens users already hold or can still create, and that half belongs to the provider. GitHub, for example, lets an organization &lt;a href="https://docs.github.com/en/organizations/managing-programmatic-access-to-your-organization/setting-a-personal-access-token-policy-for-your-organization" rel="noopener noreferrer"&gt;restrict access by personal access tokens&lt;/a&gt; and enable &lt;a href="https://docs.github.com/en/organizations/managing-oauth-access-to-your-organizations-data/about-oauth-app-access-restrictions" rel="noopener noreferrer"&gt;OAuth app access restrictions&lt;/a&gt;, so members can't authorize unapproved OAuth apps against organization resources. A provider without equivalent controls leaves the personal-token path open, and the remaining options are rotation and detection.&lt;/p&gt;

&lt;h3&gt;
  
  
  Client admission: refuse unreviewed clients at the gateway
&lt;/h3&gt;

&lt;p&gt;Client admission is the one bypass control that runs inside the gateway at connection time. The OAuth handshake is where the gateway learns which client is connecting, and with &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/access/client-registration" rel="noopener noreferrer"&gt;Client ID Metadata Documents&lt;/a&gt; (CIMD) the client's ID is an HTTPS URL on its vendor's domain. That URL is the same on every install, so the gateway can allowlist it, block it, and search for it.&lt;/p&gt;

&lt;p&gt;Speakeasy exposes this as a per-server &lt;strong&gt;Client access&lt;/strong&gt; setting with three options. &lt;strong&gt;Verified clients&lt;/strong&gt; admits a catalog of client documents Speakeasy has checked plus any URLs the organization adds, &lt;strong&gt;Any client&lt;/strong&gt; admits any valid document, and &lt;strong&gt;Off&lt;/strong&gt; admits no CIMD client. Separately, &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/secure/shadow-ai" rel="noopener noreferrer"&gt;Shadow AI&lt;/a&gt; records a decision per AI tool that enrolled devices report, and a tool marked &lt;strong&gt;Blocked&lt;/strong&gt; is refused when it authenticates, on every MCP server in the organization.&lt;/p&gt;

&lt;p&gt;Admission has real gaps, and the documentation states each one. Dynamic Client Registration (DCR) stays open to every client whatever the &lt;strong&gt;Client access&lt;/strong&gt; setting says, and a Shadow AI block takes effect only for callers that present a CIMD client ID. CIMD verifies that a document belongs to its vendor, and doesn't prove that the program on a user's machine is genuine. Admission also gates new authorizations, not existing sessions, so cutting off a client that is already connected takes a revocation. And admission only covers clients that connect to the gateway, which none of the other three bypass paths do.&lt;/p&gt;

&lt;h3&gt;
  
  
  Device-level discovery: see what runs on the laptop
&lt;/h3&gt;

&lt;p&gt;The device is the only place that can see a local stdio server. Cloudflare's post lays out the trade clearly: a hook inside the MCP client runs before the request is serialized, can deny a server not on an allowlist, and covers stdio servers that never produce network traffic, but the controls have to be reproduced in every client employees use.&lt;/p&gt;

&lt;p&gt;Two kinds of device software do this work. Hooks in the coding agent see each tool call, its server, and its arguments, which is how &lt;a href="https://www.speakeasy.com/resources/ai-agent-hooks" rel="noopener noreferrer"&gt;agent hooks&lt;/a&gt; catch calls to servers the gateway doesn't host. A device agent installed through MDM inventories the AI tools on a machine and keeps the organization's client configuration in place. Speakeasy's &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/reference/device-agent" rel="noopener noreferrer"&gt;device agent&lt;/a&gt; reapplies that configuration every 60 seconds, which is what closes the personal-account case above. It is in beta and enabled per organization. Docker takes a sandbox-side approach to the same problem: its MCP access policies can refuse registration of host-run stdio servers outright.&lt;/p&gt;

&lt;p&gt;Discovery depends on enrollment, and Speakeasy's Shadow AI documentation puts it in one line: a tool on an unenrolled machine appears nowhere. Shadow MCP detection has the same property, since sessions from an uninstrumented client produce no traffic to scan.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which bypass paths can't be closed?
&lt;/h2&gt;

&lt;p&gt;Four gaps stay at least partly open whatever an organization deploys.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Unmanaged devices.&lt;/strong&gt; A personal laptop with no device agent, no hooks, and no managed proxy is invisible to three of the four layers. On that device, the organization's only levers are credentials it holds centrally and restrictions it sets at each provider.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Providers without token controls.&lt;/strong&gt; If an upstream service lets any user mint an API key and offers admins no way to restrict it, a user can always reach that service directly. Custody at the gateway removes the convenient token but not the ability to create one.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Off-network traffic.&lt;/strong&gt; An egress rule covers traffic that passes the proxy. A device on a home network without the proxy client running sends direct MCP traffic the rule never evaluates.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tools without an MCP client.&lt;/strong&gt; Some agents call APIs directly and never speak MCP. Speakeasy's Shadow AI page notes that a decision about such a tool is recorded and never enforced, because there's no MCP authentication for the gateway to refuse.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The strongest argument against investing here follows from this list. A determined person with a shell and an API key can always reach the API, and the &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP guide&lt;/a&gt; warns that enforcing a policy before a workable alternative exists can push teams toward the bypass it's meant to stop. Both points are correct, and they argue for a narrower goal: on managed devices with organization credentials, the direct path fails, and everywhere else, a direct connection shows up in an inventory with an owner.&lt;/p&gt;

&lt;h2&gt;
  
  
  In what order should you roll out bypass controls?
&lt;/h2&gt;

&lt;p&gt;Ranked from least to most effort, with the path each step closes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Move upstream credentials into the gateway, and restrict personal tokens at providers that allow it.&lt;/strong&gt; No endpoint software, and it closes the personal-token path for every provider whose credentials move. This is realistic for every organization.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Turn on client admission.&lt;/strong&gt; Set servers to admit verified clients, and record approve or block decisions for the AI tools you already know about. It closes part of the unmanaged-client path, for clients that publish a CIMD document.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Run discovery in flag mode.&lt;/strong&gt; Deploy hooks and the device agent, and let a flag-only policy build the inventory of stdio and direct remote servers before anything is blocked. Speakeasy's &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/guides/shadow-mcp-allow-and-block-lists" rel="noopener noreferrer"&gt;shadow MCP allow and block lists guide&lt;/a&gt; describes this sequence, including the choice between a block-all and an allow-all posture.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lock down the origins you own.&lt;/strong&gt; Move internal MCP servers behind tunnels, and have them verify the gateway's signed identity header. Add private network access for the gateway endpoint if your clients all run on enrolled devices.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Block direct MCP traffic at the egress proxy.&lt;/strong&gt; This needs TLS inspection and a managed proxy on every device, which many organizations already run and many don't. If you don't have one, this step is a network rollout before it is a policy.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Each step makes the next one less disruptive. Credentials in the gateway mean a blocked direct connection has a governed alternative waiting, and an inventory built in flag mode means the first blocking policy is based on observed traffic instead of guesses.&lt;/p&gt;

&lt;h2&gt;
  
  
  A note on Speakeasy
&lt;/h2&gt;

&lt;p&gt;Speakeasy runs a managed MCP gateway as part of the AI Control Plane. Of the layers above, it covers credential custody, client admission, and origin lockdown for internal servers through tunnels, and it pairs those with device-side discovery. &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/secure/shadow-ai" rel="noopener noreferrer"&gt;Shadow AI&lt;/a&gt; is generally available under Guardrails for every organization. It lists the harnesses, assistants, and local models that enrolled devices report, and lets an admin approve or block each tool at the gateway. A blocked tool is refused when it authenticates, on every MCP server in the organization.&lt;/p&gt;

&lt;p&gt;The limits are the ones described above. Enforcement needs the tool to publish a client ID metadata document. A tool with no MCP client gets a recorded decision and no enforcement. And a device that isn't enrolled doesn't appear. Blocking direct MCP traffic on the network is a job for a secure web gateway, such as Cloudflare's, alongside the gateway.&lt;/p&gt;

&lt;p&gt;If you're mapping which of these paths are open in your organization today, we'd be interested to hear how you're approaching it. &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;Talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is MCP gateway bypass?
&lt;/h3&gt;

&lt;p&gt;MCP gateway bypass is any route by which an AI agent reaches a Model Context Protocol (MCP) server, or the API behind it, without passing through the organization's MCP gateway. The gateway's authentication, tool policy, and audit log apply only to requests it receives, so a call that takes another route carries none of them.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is MCP gateway bypass different from shadow MCP?
&lt;/h3&gt;

&lt;p&gt;Shadow MCP is a connection to an MCP server the organization has not approved. MCP gateway bypass is the broader set. It includes shadow MCP, and it also includes direct connections to approved servers that skip the gateway in front of them. Cloudflare draws the same line between shadow MCP and what it calls portal bypass, and notes that the two need different controls.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are the main ways agents bypass an MCP gateway?
&lt;/h3&gt;

&lt;p&gt;There are four common paths. An agent can launch a local stdio MCP server as a subprocess, connect directly to a remote MCP server URL, call the upstream service with a personal token the user holds, or connect through an MCP client the organization never reviewed or configured. Each path is closed by a different control layer.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can an MCP gateway prevent bypass on its own?
&lt;/h3&gt;

&lt;p&gt;No. An MCP gateway governs the requests that reach it, so it cannot block a local stdio server or a direct connection to a remote server. The gateway can close two paths itself: it can hold upstream credentials so users have no token to take elsewhere, and it can refuse unreviewed MCP clients at authentication. Network egress controls and device-level discovery close the rest.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do you stop local stdio MCP servers from bypassing the gateway?
&lt;/h3&gt;

&lt;p&gt;A local stdio server runs as a subprocess of the MCP client and sends no MCP traffic over the network, so a gateway or a network proxy never sees it. The control point is the client itself: hooks inside the coding agent can see and deny calls to servers that the gateway does not host, and a device agent can report and correct client configuration. Both require the device to be enrolled.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do you block direct connections to remote MCP servers?
&lt;/h3&gt;

&lt;p&gt;Use a network egress control on managed devices that can recognize MCP traffic and block requests that did not come through the approved gateway. Cloudflare Gateway, for example, offers an Is MCP selector and a Traffic Source selector for this. The control only applies to TLS-inspected traffic on managed network paths, so pair it with an origin that rejects requests that did not arrive through the gateway.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do personal access tokens enable MCP gateway bypass?
&lt;/h3&gt;

&lt;p&gt;A user who holds a working token for an upstream service can use it from any MCP client, script, or server, and none of those calls touch the gateway. Moving credential custody to the gateway removes that token from the user. Restricting personal tokens at the provider, where the provider supports it, stops tokens created outside the gateway from reaching organization data.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is MCP client admission?
&lt;/h3&gt;

&lt;p&gt;MCP client admission is the gateway's decision about which MCP clients may start an OAuth authorization. With Client ID Metadata Documents (CIMD), a client identifies itself by an HTTPS URL on its vendor's domain, which gives the gateway a stable identity to allowlist or block. Admission controls which clients reach the gateway; it does not see clients that never connect to the gateway.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does network egress filtering stop MCP gateway bypass?
&lt;/h3&gt;

&lt;p&gt;It stops part of it. A secure web gateway with TLS inspection can identify remote MCP traffic from managed devices and block direct connections. It cannot see local stdio servers, traffic from devices that are off the managed network, or traffic exempt from inspection, and some legacy MCP clients send their first request without the protocol header that detection relies on.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can you prevent MCP gateway bypass on unmanaged devices?
&lt;/h3&gt;

&lt;p&gt;Not reliably. Device-level discovery, client hooks, and egress filtering all depend on the device being enrolled or its traffic passing a managed proxy. On an unmanaged device, the remaining controls are the ones the organization holds centrally: keeping upstream credentials in the gateway and restricting personal tokens and unapproved OAuth apps at each provider that allows it.&lt;/p&gt;

&lt;h3&gt;
  
  
  What should you roll out first to prevent MCP gateway bypass?
&lt;/h3&gt;

&lt;p&gt;Start with credential custody, because it needs no endpoint software and closes the personal-token path for every provider whose credentials move into the gateway. Then turn on client admission, then run discovery in a flag-only mode to build an inventory, and add egress blocking last, since it needs TLS inspection and a managed proxy on every device.&lt;/p&gt;

&lt;h3&gt;
  
  
  Do agents bypass the gateway on purpose?
&lt;/h3&gt;

&lt;p&gt;Cloudflare describes the typical case: an employee finds a server in a repository, a product guide, or a colleague's message and adds it to their client without checking whether it is approved. Controls that make the governed path the default and make the direct path fail on managed devices handle that case. A determined person on an unmanaged device is a different problem and needs provider-side controls.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does a private MCP gateway prevent bypass?
&lt;/h3&gt;

&lt;p&gt;A private MCP gateway keeps the gateway's own listener off the public internet, which narrows who can reach the gateway. It does not stop an agent from reaching other servers directly. Private networking helps prevent bypass at the origin, when internal MCP servers are reachable only through the gateway, for example over an outbound-only tunnel, so a direct request has nowhere to go. On Speakeasy, private network access for the gateway endpoint runs over a Tailscale tailnet and requires an Enterprise organization.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does Speakeasy prevent MCP gateway bypass?
&lt;/h3&gt;

&lt;p&gt;Speakeasy is a managed MCP gateway. Its Shadow AI feature, generally available under Guardrails for every organization, lets admins approve or block each AI tool that enrolled devices report, and blocked tools are refused when they authenticate on every MCP server in the organization. Enforcement needs the tool to publish a client ID metadata document, and devices that are not enrolled are not visible.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>Managed vs self-hosted MCP gateway: how to choose</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Thu, 12 Nov 2026 12:42:58 +0000</pubDate>
      <link>https://dev.to/speakeasy/managed-vs-self-hosted-mcp-gateway-how-to-choose-53a9</link>
      <guid>https://dev.to/speakeasy/managed-vs-self-hosted-mcp-gateway-how-to-choose-53a9</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/managed-vs-self-hosted-mcp-gateway" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;A managed MCP gateway is a Model Context Protocol (MCP) gateway that a vendor operates on infrastructure the vendor runs. A self-hosted MCP gateway is one your team deploys, upgrades, and operates on infrastructure you control. The choice decides who carries the operations work and whose infrastructure tool-call traffic crosses. It does not, on its own, decide whether the gateway can be reached from the public internet.&lt;/p&gt;

&lt;p&gt;An &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt; sits between AI agents and the MCP servers they call. It authenticates the caller, decides which tools that caller may use, logs each call, and usually holds the upstream OAuth tokens and API keys for every system behind it. That last property is why the hosting question comes up early in security reviews. Whoever runs the gateway runs a credential store for the organization's tools.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Self-host when the data path is the requirement, meaning tool-call payloads may not cross a third party, data has to stay in a named account, or the network is air-gapped. When the requirement is that MCP traffic stays off the public internet, a managed gateway with private network access usually meets it with less work.&lt;/strong&gt; The rest of this page walks through the six questions that separate those cases.&lt;/p&gt;

&lt;h2&gt;
  
  
  The six questions that decide it
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Operations burden.&lt;/strong&gt; Who runs the database, cache, secrets, availability, and on-call for a service that holds credentials for every system behind it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Upgrade cadence.&lt;/strong&gt; Who decides when MCP specification revisions and security fixes reach production, and who absorbs the lag.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Data path and residency.&lt;/strong&gt; Whose infrastructure tool arguments and results pass through, what gets stored, where, and for how long.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Network reachability.&lt;/strong&gt; Whether agents can reach the gateway, whether the gateway can reach internal MCP servers, and whether either needs a public listener.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Identity integration.&lt;/strong&gt; Who builds and maintains the mapping from directory groups to per-tool access, including offboarding.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Compliance.&lt;/strong&gt; Whose controls and evidence the auditor reviews, and whether a contract or regulator rules out a third-party processor.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What is the difference between a managed and a self-hosted MCP gateway?
&lt;/h2&gt;

&lt;p&gt;The difference is who operates the software and on whose infrastructure the traffic runs. In practice there are several models between the two ends, and some vendors sell more than one. Ordered from least to most that your team runs:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Who operates it&lt;/th&gt;
&lt;th&gt;Where tool traffic flows&lt;/th&gt;
&lt;th&gt;Examples&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Multi-tenant managed&lt;/td&gt;
&lt;td&gt;The vendor&lt;/td&gt;
&lt;td&gt;The vendor's infrastructure&lt;/td&gt;
&lt;td&gt;Cloudflare MCP server portals, Speakeasy&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Vendor-operated in your account&lt;/td&gt;
&lt;td&gt;The vendor&lt;/td&gt;
&lt;td&gt;Your cloud account or VPC&lt;/td&gt;
&lt;td&gt;Docker MCP Enterprise Gateway, "managed in your cloud"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Managed control plane, self-hosted data plane&lt;/td&gt;
&lt;td&gt;The vendor runs configuration; you run the proxy nodes&lt;/td&gt;
&lt;td&gt;Your nodes&lt;/td&gt;
&lt;td&gt;Kong Konnect in hybrid mode&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Self-hosted&lt;/td&gt;
&lt;td&gt;Your team&lt;/td&gt;
&lt;td&gt;Your infrastructure&lt;/td&gt;
&lt;td&gt;IBM ContextForge, Docker MCP Gateway, self-managed Kong Gateway&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Air-gapped&lt;/td&gt;
&lt;td&gt;Your team, sometimes on a vendor appliance&lt;/td&gt;
&lt;td&gt;A network with no outbound connection&lt;/td&gt;
&lt;td&gt;Docker MCP Enterprise Gateway air-gapped appliance&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Docker's MCP Enterprise Gateway page lists three deployment options: Docker-operated inside your own account or VPC, an air-gapped appliance on private Kubernetes, and a multi-tenant Docker-managed cloud that the page describes as coming. Kong's deployment topologies split a cluster into a control plane, where configuration is managed, and data planes that serve traffic, and the data planes can be Kong-managed or self-managed.&lt;/p&gt;

&lt;p&gt;Exposure is a separate axis. A self-hosted gateway behind a public load balancer is as reachable from the internet as any hosted one, and a managed gateway can be private when the vendor joins it to your network. &lt;a href="https://www.speakeasy.com/resources/private-mcp-gateway" rel="noopener noreferrer"&gt;What is a private MCP gateway?&lt;/a&gt; covers that distinction in full. This page uses it only where it changes who should run the gateway.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is the case for self-hosting an MCP gateway?
&lt;/h2&gt;

&lt;p&gt;The case for self-hosting is strong, and for some organizations it settles the question before any other criterion is considered:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The gateway sees the most sensitive traffic agents produce.&lt;/strong&gt; Tool arguments and results carry customer records, source code, tickets, and query results. Running the gateway yourself removes a processor from that data flow and a vendor from the list of parties whose breach becomes yours.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Upgrades happen on your schedule.&lt;/strong&gt; You can test each release against your own MCP servers, pin versions in change management, and roll back without waiting on anyone.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You can read the code.&lt;/strong&gt; IBM ContextForge is Apache 2.0 licensed and Docker MCP Gateway is MIT licensed. A security team can audit what the gateway does with credentials instead of relying on a vendor's description.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Internal servers are already reachable.&lt;/strong&gt; A gateway deployed in the same VPC as your internal MCP servers needs no tunnel and no connector.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Some environments allow nothing else.&lt;/strong&gt; Air-gapped networks, and contracts that forbid third-party processing, rule out multi-tenant managed gateways outright.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each of these is a real advantage, but the cost is that the gateway becomes a tier-zero service your team runs, upgrades, and answers pages for. And several requirements that start the self-hosting conversation turn out, when written down, to be about exposure, which both models can meet.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you decide between a managed and a self-hosted MCP gateway?
&lt;/h2&gt;

&lt;p&gt;Work through the six questions in order. Data path and reachability are the two that can rule an option out entirely, while operations, upgrades, identity, and compliance mostly change what each option costs.&lt;/p&gt;

&lt;h3&gt;
  
  
  Operations burden: who runs a tier-zero service?
&lt;/h3&gt;

&lt;p&gt;A self-hosted MCP gateway is more than a proxy process. IBM ContextForge's documentation recommends PostgreSQL and Redis alongside the gateway for production, offers Docker Compose and Helm deployments, points to a configuration reference of more than 300 environment variables, and lists secrets that must be set before production, including the JWT signing key, the credential encryption key, and the bootstrap admin password. Any gateway that stores credentials needs the same.&lt;/p&gt;

&lt;p&gt;Around the gateway itself, your team takes on high availability, backups of the token store, rotation of the key that encrypts stored credentials, TLS, monitoring, log export, base image patching, and on-call. Multi-node deployments need extra care with session state. Kong's &lt;a href="https://developer.konghq.com/plugins/ai-mcp-proxy/#scope-of-support" rel="noopener noreferrer"&gt;AI MCP Proxy plugin&lt;/a&gt; lists session sharing between instances as not supported, which is worth testing before putting several data plane nodes behind one load balancer.&lt;/p&gt;

&lt;p&gt;A managed gateway moves that work to the vendor and replaces it with a dependency. When the vendor's platform is down, the MCP servers behind it are unreachable, and your incident response is a status page. Kong's hybrid mode reduces this risk for the traffic path, because its data planes keep serving with their last known configuration when the control plane is offline.&lt;/p&gt;

&lt;p&gt;The deciding question is whether you already run a platform team that operates stateful, credential-holding services with on-call. If you do, a self-hosted gateway is one more of those. If you don't, the gateway would be the first, and standing up that capability is a larger project than the gateway.&lt;/p&gt;

&lt;h3&gt;
  
  
  Upgrade cadence: who decides when the gateway changes?
&lt;/h3&gt;

&lt;p&gt;MCP changes at the protocol level often enough that gateway versions matter. Kong's AI MCP Proxy documentation says the plugin uses protocol version 2025-06-18, supports upstream servers on 2025-06-18 and 2025-11-25, and doesn't support 2024 versions. It also records that before Kong Gateway 3.14, a request denied by an access rule returned a JSON-RPC &lt;code&gt;INVALID_PARAMS&lt;/code&gt; error, and that 3.14 changed this to an HTTP &lt;code&gt;403&lt;/code&gt; to match the 2025-11-25 authorization specification. Cloudflare's MCP server portals already accept stateless MCP &lt;code&gt;2026-07-28&lt;/code&gt; clients alongside 2025 Streamable HTTP. ContextForge introduced multi-tenancy, teams, and RBAC in v0.7.0, and its SSO documentation tells deployments upgrading from 0.6.0 to review the migration guide and changelog first.&lt;/p&gt;

&lt;p&gt;Every one of those changes is something an operator has to roll out. With a self-hosted gateway, you choose the timing and test each release against your servers first. You also own the lag: when a client adopts a newer protocol revision, or a security fix ships, it waits for your next upgrade window. With a managed gateway, changes arrive when the vendor ships them. You don't schedule the work, and you can't defer a change that breaks one of your workflows, so ask each vendor how it announces breaking changes and whether changes can be staged.&lt;/p&gt;

&lt;p&gt;The deciding question is how your change management treats infrastructure. If every version is pinned and tested before production, self-hosting fits the process you already have. If your process accepts SaaS changes under a vendor's notice policy, a managed gateway removes the upgrade work entirely.&lt;/p&gt;

&lt;h3&gt;
  
  
  Data path and residency: whose infrastructure sees tool traffic?
&lt;/h3&gt;

&lt;p&gt;This is the question that can rule a managed gateway out. Every call through an MCP gateway carries the tool's arguments and its result, and with a multi-tenant managed gateway both cross the vendor's infrastructure. Anything the gateway stores, such as logs, transcripts, or upstream tokens, lives there too.&lt;/p&gt;

&lt;p&gt;With a managed vendor, ask:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;What does the gateway store by default, and what is opt-in?&lt;/li&gt;
&lt;li&gt;How long is stored data retained, and can you shorten it?&lt;/li&gt;
&lt;li&gt;In which regions is data processed and stored, and can you choose?&lt;/li&gt;
&lt;li&gt;Which subprocessors handle sign-in, logging, or storage?&lt;/li&gt;
&lt;li&gt;Who controls the keys that encrypt stored credentials and sign tokens?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A self-hosted gateway keeps traffic and storage wherever you deploy it, and it makes you the custodian of every upstream credential it brokers. ContextForge, for example, requires an &lt;code&gt;AUTH_ENCRYPTION_SECRET&lt;/code&gt; passphrase for encrypting stored credentials, so generating, storing, and rotating that key is your team's job.&lt;/p&gt;

&lt;p&gt;Between the two, Docker's vendor-operated option runs the gateway inside your own account or VPC, where, per Docker, data and traffic stay in your environment. For an air-gapped network, Docker's appliance runs on private Kubernetes with no outbound connection. A multi-tenant managed gateway cannot meet an air-gap requirement under any configuration.&lt;/p&gt;

&lt;p&gt;The deciding question is whether tool arguments and results may cross a third party's infrastructure under a data processing agreement. If they may not, multi-tenant managed gateways are out, and the choice is between self-hosting and a vendor-operated deployment in your account. If they may, the data path becomes an item in vendor review rather than an architecture decision.&lt;/p&gt;

&lt;h3&gt;
  
  
  Network reachability: what has to be public?
&lt;/h3&gt;

&lt;p&gt;Reachability runs in two directions, and the two favor different models.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;From the gateway to MCP servers.&lt;/strong&gt; A self-hosted gateway in your VPC reaches internal MCP servers directly. A managed gateway needs a route in, usually an &lt;a href="https://www.speakeasy.com/resources/mcp-tunnels" rel="noopener noreferrer"&gt;MCP tunnel&lt;/a&gt;: a small agent inside your network opens an outbound connection to the gateway, so internal servers need no public address and no inbound firewall rule. Check the details per vendor. Cloudflare's portals reach private MCP servers through Cloudflare Tunnel with Gateway routing turned on, but the &lt;a href="https://developers.cloudflare.com/cloudflare-one/access-controls/ai-controls/mcp-portals/#connect-a-private-mcp-server" rel="noopener noreferrer"&gt;documentation&lt;/a&gt; states that the OAuth authorization and token endpoints must be accessible on the public internet.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;From agents to the gateway.&lt;/strong&gt; List your callers. Agents on managed laptops and CI runners can sit on a private network. Hosted agents, such as remote connectors in ChatGPT or Claude, call MCP servers from the vendor's cloud and can only reach a public endpoint. That holds regardless of who runs the gateway. A self-hosted gateway that has to serve hosted agents still needs public ingress, and its exposure is then whatever your load balancer, ingress controller, and DNS say it is, which can change in a later deploy.&lt;/p&gt;

&lt;p&gt;The requirement that often sits behind "we need to self-host" is "the gateway can't have a public listener." That is a reachability requirement, and a managed gateway meets it if the vendor can join the gateway to your private network. Few managed MCP gateways document this. Where one does, you keep vendor operations and remove the public listener, at the costs the &lt;a href="https://www.speakeasy.com/resources/private-mcp-gateway" rel="noopener noreferrer"&gt;private MCP gateway&lt;/a&gt; page lays out: every client has to be enrolled in the network, a network outage takes tool access with it, and OAuth flows to an external identity provider usually still need the internet. Speakeasy offers this as Tailscale private access for Enterprise organizations, covered in the note at the end of this page.&lt;/p&gt;

&lt;p&gt;The deciding question is which of your callers run on devices you manage, which run in a vendor's cloud, and which of your MCP servers live inside your network. If hosted agents are callers, some path has to be public whichever model you choose. If every caller is on your network, both a self-hosted gateway in your VPC and a managed gateway with private access work, and the choice falls back to operations and data path.&lt;/p&gt;

&lt;h3&gt;
  
  
  Identity integration: who maintains the mapping?
&lt;/h3&gt;

&lt;p&gt;Both models integrate with your identity provider. The difference is who builds the mapping from directory groups to per-tool permissions and who keeps it working. On the self-hosted side, ContextForge documents SSO with GitHub, Google, IBM Security Verify, Microsoft Entra ID, Keycloak, Okta, and generic OIDC, creates users on their first SSO login, and maps identity provider groups to teams. Kong authenticates MCP callers with its OpenID Connect or Key Auth plugins and, from version 3.13, applies per-tool ACLs to Consumers and Consumer Groups.&lt;/p&gt;

&lt;p&gt;Whichever model you choose, check three things:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Offboarding.&lt;/strong&gt; Creating a user on first login handles joiners. Removing access when someone leaves the directory needs SCIM or another sync, so confirm which one the gateway supports and test it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Upstream credentials.&lt;/strong&gt; Decide whether each user connects their own account to each upstream service, or whether one shared credential per server is acceptable, and who can revoke either.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The identity provider's own location.&lt;/strong&gt; Okta and Microsoft Entra ID are SaaS, so self-hosting the gateway rarely removes the internet dependency for sign-in.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The deciding question is whether your identity team wants to own the group-to-tool mapping as configuration it maintains, which suits a self-hosted gateway, or wants directory sync with a vendor keeping the integration current, which suits a managed one.&lt;/p&gt;

&lt;h3&gt;
  
  
  Compliance: whose controls get audited?
&lt;/h3&gt;

&lt;p&gt;Compliance frameworks ask for controls and evidence: access control, logging, change management, encryption, and vendor management. They rarely specify who runs the software. With a managed gateway, the vendor's SOC 2 Type II or ISO 27001 report covers its operation of the gateway, you add the vendor to your third-party risk program, and your evidence covers your own configuration. With a self-hosted gateway, the gateway falls inside your audit scope, so your team produces the evidence for patching, access reviews, log retention, and key management.&lt;/p&gt;

&lt;p&gt;Neither model makes an organization compliant. Both need the same per-call audit trail exported to your SIEM, which self-hosted options support through OpenTelemetry in ContextForge and an audit log of allowed and denied ACL decisions in Kong's plugin. Where the two diverge is a contract or regulator that forbids a third-party processor, or a government network that requires an authorization the vendor doesn't hold. Those rule out multi-tenant managed gateways the same way the data path question does.&lt;/p&gt;

&lt;p&gt;The deciding question is whether your auditor will accept a vendor's report for the gateway layer. If they will, a managed gateway reduces your evidence work. If a contract rules out the processor, self-host or use a vendor-operated deployment in your account.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which should you choose?
&lt;/h2&gt;

&lt;p&gt;The table summarizes where each question points.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Question&lt;/th&gt;
&lt;th&gt;Points to managed when&lt;/th&gt;
&lt;th&gt;Points to self-hosted when&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Operations burden&lt;/td&gt;
&lt;td&gt;No team runs stateful, credential-holding services with on-call&lt;/td&gt;
&lt;td&gt;A platform team already operates services like it&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Upgrade cadence&lt;/td&gt;
&lt;td&gt;Vendor-scheduled changes with notice are acceptable&lt;/td&gt;
&lt;td&gt;Every version has to be pinned and tested first&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Data path and residency&lt;/td&gt;
&lt;td&gt;Tool traffic may cross a vendor under a data processing agreement&lt;/td&gt;
&lt;td&gt;Tool traffic may not leave your account, region, or network&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Network reachability&lt;/td&gt;
&lt;td&gt;Hosted agents are callers, or the vendor can join your private network&lt;/td&gt;
&lt;td&gt;Every caller and server is already on a network you run&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Identity integration&lt;/td&gt;
&lt;td&gt;You want directory sync maintained by the vendor&lt;/td&gt;
&lt;td&gt;Your identity team wants to own the mapping as configuration&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Compliance&lt;/td&gt;
&lt;td&gt;The auditor accepts a vendor's SOC 2 or ISO 27001 report&lt;/td&gt;
&lt;td&gt;A contract or regulator rules out a third-party processor&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Each choice also has its own way of going wrong. A managed gateway can go down with its vendor, change under you on the vendor's schedule, or support only the private networks the vendor has integrated with. A self-hosted gateway can fall behind on upgrades until it is the least-patched service holding the most credentials, drift into public exposure after a load balancer or DNS change, and turn a credential-store breach into your incident rather than a vendor's. Picking managed accepts the first set of risks, and picking self-hosted accepts the second.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which self-hosted MCP gateways are available?
&lt;/h2&gt;

&lt;p&gt;Each entry covers what the vendor documents as of October 2026. &lt;a href="https://www.speakeasy.com/blog/best-mcp-gateways-for-enterprise-2026" rel="noopener noreferrer"&gt;The best MCP gateways for enterprise in 2026&lt;/a&gt; scores a wider set of products, and &lt;a href="https://www.speakeasy.com/blog/choosing-an-mcp-gateway" rel="noopener noreferrer"&gt;Which MCP gateway architecture do I need?&lt;/a&gt; compares routing infrastructure with governance platforms. For head-to-head comparisons, see &lt;a href="https://www.speakeasy.com/blog/speakeasy-vs-docker-mcp-gateway" rel="noopener noreferrer"&gt;Speakeasy vs Docker MCP Gateway&lt;/a&gt;, &lt;a href="https://www.speakeasy.com/blog/speakeasy-vs-ibm-contextforge" rel="noopener noreferrer"&gt;Speakeasy vs IBM ContextForge&lt;/a&gt;, &lt;a href="https://www.speakeasy.com/blog/speakeasy-vs-kong-mcp-gateway" rel="noopener noreferrer"&gt;Speakeasy vs Kong&lt;/a&gt;, and &lt;a href="https://www.speakeasy.com/blog/speakeasy-vs-cloudflare-mcp-server-portals" rel="noopener noreferrer"&gt;Speakeasy vs Cloudflare MCP server portals&lt;/a&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Docker MCP Gateway
&lt;/h3&gt;

&lt;p&gt;Docker MCP Gateway is Docker's open source gateway. It runs each MCP server in an isolated container with restricted privileges, network access, and resources, starts servers on demand, injects credentials, and includes logging and call tracing. It powers the MCP Toolkit in Docker Desktop and can run on Docker Engine without Desktop, with servers grouped into profiles and per-tool allowlists set from the CLI. Its strength is server isolation, which matters most for the local npx and uvx servers that otherwise run with a developer's full permissions.&lt;/p&gt;

&lt;p&gt;The open source gateway is built around a developer's machine. Organization-wide governance, with SSO, group-scoped tools, per-call policy, and SIEM audit, is in the separate MCP Enterprise Gateway, which Docker sells through its sales team, and Docker's documentation describes MCP Gateway as part of Docker AI Governance as invite-only. The Enterprise Gateway's deployment options (in your account, air-gapped, and a planned Docker-managed cloud) are in the table near the top of this page.&lt;/p&gt;

&lt;h3&gt;
  
  
  IBM ContextForge
&lt;/h3&gt;

&lt;p&gt;ContextForge is an open source registry and proxy that federates MCP servers, A2A agents, and REST or gRPC APIs behind one endpoint. It adds virtual servers that bundle selected tools, an admin UI, more than 40 plugins, OpenTelemetry tracing, and multi-cluster Kubernetes deployment with Redis-backed federation, and it supports SSO with the major identity providers. It suits a platform team that wants to self-host and bring existing non-MCP services into the same catalog.&lt;/p&gt;

&lt;p&gt;The trade is operational scope: production use means running ContextForge with PostgreSQL and Redis, working through a large configuration surface, and following its migration guides between versions, as the operations and upgrade sections above describe.&lt;/p&gt;

&lt;h3&gt;
  
  
  Kong AI Gateway
&lt;/h3&gt;

&lt;p&gt;Kong adds MCP to its API gateway through the AI MCP Proxy plugin, which requires the AI Gateway Enterprise offering. Kong AI Gateway 2.0 replaces the plugin with an AI MCP Server entity, so the details here are for the 3.x plugin. The plugin can proxy existing MCP servers, convert REST APIs described by OpenAPI into MCP tools, and aggregate tools from several services behind one MCP endpoint. Kong's authentication, rate limiting, and logging plugins then apply to MCP traffic, and per-tool ACLs filter both tool listing and invocation. For an organization that already runs Kong for its APIs, this puts MCP under the policies and operations it already has, and hybrid mode lets Kong run the control plane while you host the data planes.&lt;/p&gt;

&lt;p&gt;Kong documents its gaps plainly. The plugin's scope-of-support table lists AI Guardrails on MCP requests and responses, WebSocket and gRPC upstreams, structured output, tool-change notifications, and session sharing between instances as not supported.&lt;/p&gt;

&lt;h3&gt;
  
  
  Managed options to compare against
&lt;/h3&gt;

&lt;p&gt;Cloudflare MCP server portals, part of Cloudflare One, are the closest managed comparison for an organization already on Cloudflare. A portal puts several MCP servers behind one endpoint, uses Cloudflare Access identity policies to decide who reaches it, curates and renames tools, logs tool requests, and can route traffic through Cloudflare Gateway for HTTP logging and DLP scanning. Private servers connect over Cloudflare Tunnel. Cloudflare's documentation notes two boundaries: OAuth authorization and token endpoints for private servers must be public, and a user blocked by a portal's Access policies can still &lt;a href="https://www.speakeasy.com/resources/mcp-gateway-bypass" rel="noopener noreferrer"&gt;reach an MCP server through its direct URL&lt;/a&gt; unless Access is configured as that server's OAuth provider.&lt;/p&gt;

&lt;h2&gt;
  
  
  A note on Speakeasy
&lt;/h2&gt;

&lt;p&gt;We build the Speakeasy MCP gateway, so this section isn't neutral. The gateway runs as part of the &lt;a href="https://www.speakeasy.com/resources/ai-control-plane" rel="noopener noreferrer"&gt;AI Control Plane&lt;/a&gt; and is managed by Speakeasy by default.&lt;/p&gt;

&lt;h3&gt;
  
  
  Need to self-host?
&lt;/h3&gt;

&lt;p&gt;Self-hosting is available for Enterprise customers. &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;Book time with our team&lt;/a&gt; to learn more.&lt;/p&gt;

&lt;p&gt;For teams whose requirement is reachability, Speakeasy's &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/private-network-access/tailscale" rel="noopener noreferrer"&gt;Tailscale private access&lt;/a&gt; is the middle path, and it requires an Enterprise organization. An org admin connects a tailnet once with a dedicated Tailscale OAuth client, and the gateway gets a private hostname on that tailnet. Each MCP server or gateway is then set to public only, public and private, or private only, so clients can move one server at a time. Private-only servers stop serving on public URLs with no automatic public fallback, so a tailnet outage leaves them unreachable rather than exposed. Tailscale membership adds a network check, and MCP authentication and tool permissions still apply to every private connection.&lt;/p&gt;

&lt;p&gt;The limits are specific. Tailscale is the only private network provider today. Private access covers how MCP clients reach hosted endpoints, and it doesn't make the dashboard, management APIs, or outbound connections to tool backends private. For servers inside your network, &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/tunneled-servers" rel="noopener noreferrer"&gt;tunneled MCP servers&lt;/a&gt; run a tunnel agent that opens one outbound connection with no inbound firewall rule, and the two can be combined so neither end of the path is public. The &lt;a href="https://www.speakeasy.com/blog/release-private-mcp-gateway-tailscale" rel="noopener noreferrer"&gt;private MCP gateway release post&lt;/a&gt; describes that pairing.&lt;/p&gt;

&lt;p&gt;On the other four questions, Speakeasy ships changes to the hosted gateway continuously and records them in the &lt;a href="https://www.speakeasy.com/changelog" rel="noopener noreferrer"&gt;changelog&lt;/a&gt;. Its &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/reference/platform-security" rel="noopener noreferrer"&gt;platform security documentation&lt;/a&gt; says the platform processes customer data in transit, makes storage of session transcripts, tool telemetry, and tool inputs and outputs opt-in, and retains stored logs for 90 days by default. It names WorkOS as the dependency for SSO and directory sync, and it doesn't describe a customer-selected data region, so ask if residency in a named region is contractual. &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/org-admin/encryption-keys" rel="noopener noreferrer"&gt;Customer-managed signing keys&lt;/a&gt; in your own cloud KMS are enabled per organization. &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/org-admin/identity" rel="noopener noreferrer"&gt;Directory Sync&lt;/a&gt; assigns roles from identity provider groups and removes access when people leave the directory, &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/org-admin/roles-and-permissions" rel="noopener noreferrer"&gt;roles&lt;/a&gt; narrow access to specific servers or tools, and &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/identity/identities/enterprise-managed-auth" rel="noopener noreferrer"&gt;Enterprise Managed Auth&lt;/a&gt; lets people sign in once with Okta and reach their MCP servers without a separate OAuth prompt per service. Administrative changes appear in &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/org-admin/audit-logs" rel="noopener noreferrer"&gt;Audit Logs&lt;/a&gt;, and Speakeasy holds SOC 2 Type II and ISO 27001 (&lt;a href="https://www.speakeasy.com/security" rel="noopener noreferrer"&gt;security&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;If you're partway through a security review and the self-hosting question is still open, we'd like to hear which requirement is driving it. &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;Talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is the difference between a managed and a self-hosted MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A managed MCP gateway is a Model Context Protocol (MCP) gateway that a vendor operates on infrastructure the vendor runs. A self-hosted MCP gateway is one your team deploys, upgrades, and operates on infrastructure you control. The choice decides who carries the operations work and whose infrastructure tool-call traffic crosses. It does not, on its own, decide whether the gateway can be reached from the public internet.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is a self-hosted MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A self-hosted MCP gateway is an MCP gateway your team deploys and runs on your own servers, Kubernetes clusters, or cloud accounts. Your team owns its database, upgrades, availability, secrets, and on-call, and tool-call traffic stays on infrastructure you control. Open source examples include IBM ContextForge and Docker MCP Gateway, and Kong Gateway can be self-managed with the AI MCP Proxy plugin.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is a managed MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A managed MCP gateway is an MCP gateway a vendor operates for you, usually as a multi-tenant cloud service. The vendor runs the infrastructure, ships upgrades, and handles availability, while your team configures servers, access, and policy. Tool-call arguments and results pass through the vendor's infrastructure, so the vendor becomes a processor in your data flow. Examples include Cloudflare MCP server portals and the Speakeasy MCP gateway.&lt;/p&gt;

&lt;h3&gt;
  
  
  When should you self-host an MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Self-host an MCP gateway when the data path is the requirement: tool-call arguments and results may not cross a third party's infrastructure, data has to stay in a named account or region, or the network is air-gapped. Self-hosting also fits a platform team that already operates stateful, credential-holding services with on-call and wants to pin and test every version. If the requirement is only that MCP traffic stays off the public internet, a managed gateway with private network access can meet it.&lt;/p&gt;

&lt;h3&gt;
  
  
  When is a managed MCP gateway the better choice?
&lt;/h3&gt;

&lt;p&gt;A managed MCP gateway is the better choice when tool traffic can pass through a vendor under a data processing agreement and your team would rather not run a tier-zero service. The vendor handles upgrades as the MCP specification changes, availability, and the credential store's infrastructure. Your audit then relies partly on the vendor's SOC 2 or ISO 27001 report, and your availability depends on theirs.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is a self-hosted MCP gateway more secure than a managed one?
&lt;/h3&gt;

&lt;p&gt;Not by default. A self-hosted MCP gateway removes a third party from the data path, but it also makes your team responsible for patching, secrets, the credential store, and exposure. A gateway that falls behind on upgrades or sits behind a misconfigured public load balancer is a larger risk than a well-run managed one. Security depends on how well the gateway is operated, whoever operates it.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is a self-hosted MCP gateway the same as a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;No. Self-hosted describes who operates the gateway; private describes who can reach it. A self-hosted MCP gateway behind a public load balancer is exposed to the internet, and a managed MCP gateway can be private if the vendor joins it to your private network. Many requirements phrased as self-hosting are exposure requirements that a private MCP gateway meets.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can a managed MCP gateway reach MCP servers inside a private network?
&lt;/h3&gt;

&lt;p&gt;Yes, through an MCP tunnel or a private network connector. A tunnel agent inside your network opens an outbound connection to the managed gateway, so internal MCP servers get a route without an inbound firewall rule or a public address. Cloudflare MCP server portals reach private servers over Cloudflare Tunnel, and Speakeasy uses tunneled MCP servers. Check whether the vendor also requires any OAuth endpoints to stay public.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can ChatGPT or Claude connect to a self-hosted MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Hosted agents such as remote connectors in ChatGPT and Claude call MCP servers from the vendor's cloud, so they can reach a self-hosted MCP gateway only if it has a public endpoint. Agents that run on managed devices, such as Claude Code or Cursor on an employee laptop, can reach a gateway on a private network. Self-hosting doesn't change this: if hosted agents are callers, some path to the gateway has to be public.&lt;/p&gt;

&lt;h3&gt;
  
  
  What does it take to operate a self-hosted MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Operating a self-hosted MCP gateway means running the gateway plus its database and cache, high availability, backups of the token store, TLS, encryption key management, monitoring, log export, patching, and on-call. IBM ContextForge, for example, recommends PostgreSQL and Redis for production and documents more than 300 configuration variables. Because the gateway holds upstream credentials for every system behind it, it has to be run as tier-zero infrastructure.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do MCP specification updates affect a self-hosted gateway?
&lt;/h3&gt;

&lt;p&gt;Each MCP specification revision is a change someone has to roll out. Kong's AI MCP Proxy plugin documents which protocol versions it supports and changed its access-denied response in Kong Gateway 3.14 to match the 2025-11-25 authorization specification. With a self-hosted gateway you choose when to take each release and test it first, and you also own the lag when clients move to a newer revision before you upgrade.&lt;/p&gt;

&lt;h3&gt;
  
  
  Which open source MCP gateways can you self-host?
&lt;/h3&gt;

&lt;p&gt;IBM ContextForge is an Apache 2.0 licensed registry and proxy that federates MCP, A2A, and REST or gRPC services, deployable with Docker, PyPI, or Helm. Docker MCP Gateway is an MIT licensed gateway that runs each MCP server in its own container and powers the MCP Toolkit in Docker Desktop. Kong Gateway can be self-managed, though its AI MCP Proxy plugin requires the AI Gateway Enterprise offering.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is there a middle ground between managed and self-hosted MCP gateways?
&lt;/h3&gt;

&lt;p&gt;Yes, several. Docker offers its MCP Enterprise Gateway operated by Docker inside your own cloud account or VPC. Kong's hybrid mode runs the control plane in Konnect while you host the data plane nodes that carry traffic. A managed gateway with private network access, such as Speakeasy's Tailscale private access on Enterprise organizations, keeps vendor operations while removing the public listener.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can an MCP gateway run air-gapped?
&lt;/h3&gt;

&lt;p&gt;Some can. Docker documents an air-gapped appliance for its MCP Enterprise Gateway that runs on private Kubernetes with no outbound connection, loading catalogs, server images, and policy from inside the network. A self-hosted open source gateway can also run on an isolated network if its dependencies are mirrored internally. Multi-tenant managed MCP gateways cannot run air-gapped.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is a self-hosted MCP server the same as a self-hosted MCP gateway?
&lt;/h3&gt;

&lt;p&gt;No. Hosting an MCP server, yourself or with a provider, runs the server itself and gives it a remote URL. An MCP gateway sits in front of many MCP servers, hosted or not, and authenticates callers, enforces per-tool policy, brokers upstream credentials, and logs each call. Many managed platforms do both, and the managed vs self-hosted question applies to each layer separately.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does Speakeasy offer a self-hosted MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Yes, for Enterprise customers. The Speakeasy MCP gateway runs as a managed service by default, as part of the Speakeasy AI Control Plane, and self-hosting is available on Enterprise; talk to the Speakeasy team to learn more. For teams that need private reachability without self-hosting, Tailscale private access on Enterprise organizations puts the gateway on your tailnet with each server set to public only, public and private, or private only, and tunneled MCP servers reach internal servers without an inbound port.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What is a private MCP gateway?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Thu, 12 Nov 2026 11:47:14 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-a-private-mcp-gateway-32k4</link>
      <guid>https://dev.to/speakeasy/what-is-a-private-mcp-gateway-32k4</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/private-mcp-gateway" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Private MCP gateway
&lt;/h2&gt;

&lt;p&gt;A private MCP gateway is a Model Context Protocol (MCP) gateway whose listener&lt;br&gt;
is reachable only from a private network, such as a VPN, a Tailscale tailnet,&lt;br&gt;
or a cloud VPC. Agents reach it over that network, and the public internet has&lt;br&gt;
no path to the MCP servers behind it.&lt;/p&gt;

&lt;p&gt;An &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt; is the one component every governed agent tool call passes through, which makes its exposure the exposure of every MCP server behind it. Most MCP gateways, hosted and self-hosted alike, listen on the public internet and rely on OAuth, and sometimes an IP allowlist, to keep strangers out. For many teams that's a reasonable default. Organizations that already route employee traffic through a VPN or an overlay network have started asking why agent tool traffic should be the exception.&lt;/p&gt;

&lt;p&gt;Anything that speaks MCP on a public address gets found. In 2025, Knostic used Shodan to &lt;a href="https://www.knostic.ai/blog/mapping-mcp-servers-study" rel="noopener noreferrer"&gt;identify 1,862 MCP servers exposed to the internet&lt;/a&gt;, and every one of the 119 it checked by hand listed its tools to an unauthenticated caller. Those were individual servers with no authentication, and a gateway that enforces OAuth is in a much stronger position. The study still shows how quickly a public MCP endpoint turns up in someone else's scan results.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where a gateway runs and who can reach it are separate decisions, and a private MCP gateway is defined entirely by the second.&lt;/strong&gt; It says nothing about who operates the software or which building the hardware sits in, which is why it's easy to confuse with self-hosted, on-prem, and tunneled deployments.&lt;/p&gt;

&lt;h2&gt;
  
  
  The properties of a private MCP gateway
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;No public path to MCP servers.&lt;/strong&gt; The gateway listens on a private hostname or address. Public hostnames either don't exist or refuse every request.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reachability set by network policy.&lt;/strong&gt; Whether a person, device, or workload can reach the gateway is decided by the private network's access rules, such as tailnet ACLs, security groups, or VPN policy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Network identity on every request.&lt;/strong&gt; Overlay networks such as Tailscale can tell the gateway which user and device a connection came from, in addition to the OAuth identity the MCP client presents.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fails closed.&lt;/strong&gt; A well-built private gateway becomes unreachable when the private network is down or misconfigured, instead of falling back to a public route.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Exposure as an explicit setting.&lt;/strong&gt; Implementations that support both public and private access make exposure a per-gateway or per-server choice, so teams can move servers one at a time.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What is a private MCP gateway?
&lt;/h2&gt;

&lt;p&gt;A private MCP gateway is an MCP gateway placed on a private network. It does the usual gateway job: it authenticates the MCP client, decides which tools the caller may use, and logs each call. The difference is where its listener sits. Because the gateway is only reachable inside the private network, the first check on any request is whether the caller can reach the gateway at all.&lt;/p&gt;

&lt;p&gt;That produces two layers of policy, usually with different owners:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The network layer&lt;/strong&gt; decides who can open a connection, based on device enrollment and network access rules. The network or IT team typically owns it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The gateway layer&lt;/strong&gt; decides what a connected caller may do, based on OAuth identity and per-tool permissions. The platform or security team typically owns it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A public MCP gateway has only the second layer. A request to a private MCP gateway follows this path:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The gateway joins the private network and gets a private hostname. It has no public address, or its public hostnames refuse every request.&lt;/li&gt;
&lt;li&gt;An agent on an enrolled device, such as a developer laptop or a CI runner, connects to that hostname over the private network. The network admits the device before any traffic reaches the gateway.&lt;/li&gt;
&lt;li&gt;The gateway authenticates the MCP client, attaches the network identity where the network provides one, and checks policy for the tool being called.&lt;/li&gt;
&lt;li&gt;The gateway forwards the call to an internal MCP server and writes it to the audit log.&lt;/li&gt;
&lt;li&gt;A request from the public internet has nothing to connect to.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/private-mcp-gateway#:~:text=A%20request%20from%20the%20public%20internet%20has%20nothing%20to%20connect%20to" rel="noopener noreferrer"&gt;View the private MCP gateway flow diagram on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The strongest objection is that the gateway already authenticates every request, so hiding it adds little. OAuth 2.1 with PKCE is a strong scheme, and a public gateway that enforces it correctly is not an open door. But authentication runs after the gateway accepts a connection. A public listener exposes the TLS stack, the OAuth authorization and token endpoints, the protected resource metadata the MCP authorization specification requires servers to publish, and any request parsing that happens before a token is checked. A private listener limits that surface to devices the network has already admitted.&lt;/p&gt;

&lt;p&gt;The trade is uneven. Moving the listener costs a network change and some client enrollment. A pre-authentication bug in a public gateway is an incident for every MCP server behind it.&lt;/p&gt;

&lt;h2&gt;
  
  
  How is a private MCP gateway different from a self-hosted one?
&lt;/h2&gt;

&lt;p&gt;Self-hosted describes who operates the gateway. Private describes who can reach it. The two vary independently, and all four combinations exist:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Public listener&lt;/th&gt;
&lt;th&gt;Private listener only&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Managed&lt;/strong&gt; (a vendor runs the gateway)&lt;/td&gt;
&lt;td&gt;The default for most hosted MCP gateways&lt;/td&gt;
&lt;td&gt;A managed gateway joined to the customer's private network, such as Speakeasy on a Tailscale tailnet&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Self-hosted&lt;/strong&gt; (you run the gateway)&lt;/td&gt;
&lt;td&gt;A self-hosted gateway behind a public load balancer or ingress&lt;/td&gt;
&lt;td&gt;A self-hosted gateway behind an internal load balancer, reached over a VPN or overlay&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A self-hosted gateway on a public load balancer is as exposed as any hosted one. A managed gateway can be private when the vendor puts it on the customer's network. The two terms get conflated because the most common way self-hosters end up private is by deploying into a VPC with no public ingress, where the gateway is private by default. That privacy belongs to the deployment. A later change to the load balancer, the ingress controller, or DNS can make the same self-hosted gateway public again, which is why the exposure needs to be tested, not assumed. Enterprise buyers often phrase an exposure requirement as a self-hosting one, and the &lt;a href="https://www.speakeasy.com/resources/enterprise-mcp-gateway" rel="noopener noreferrer"&gt;enterprise MCP gateway guide&lt;/a&gt; explains how to separate the two.&lt;/p&gt;

&lt;p&gt;An on-prem MCP gateway runs on hardware in the organization's own facilities. That is a statement about location, and on-prem networks publish services to the internet all the time. On-prem gateways are often private in practice because the surrounding network is, but the posture still has to be configured and checked the same way.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is a private MCP gateway the same as an MCP tunnel?
&lt;/h2&gt;

&lt;p&gt;A private MCP gateway and an MCP tunnel cover opposite ends of the same path. An &lt;a href="https://www.speakeasy.com/resources/mcp-tunnels" rel="noopener noreferrer"&gt;MCP tunnel&lt;/a&gt; is an outbound-only connection that lets a gateway or agent platform reach an MCP server inside a private network without opening an inbound port. It makes a private server reachable. A private MCP gateway keeps the gateway's own listener off the public internet, so the path from agents to the gateway is private.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;What stays off the public internet&lt;/th&gt;
&lt;th&gt;Direction of the connection&lt;/th&gt;
&lt;th&gt;When you'd use it&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;MCP tunnel&lt;/td&gt;
&lt;td&gt;The MCP server&lt;/td&gt;
&lt;td&gt;Outbound, from the server's network to the gateway or agent platform&lt;/td&gt;
&lt;td&gt;A hosted gateway or agent needs an internal MCP server&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Private MCP gateway&lt;/td&gt;
&lt;td&gt;The gateway's listener&lt;/td&gt;
&lt;td&gt;Inbound, from agents over the private network to the gateway&lt;/td&gt;
&lt;td&gt;Agents on company devices call tools with no public ingress&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Both together&lt;/td&gt;
&lt;td&gt;The server and the gateway&lt;/td&gt;
&lt;td&gt;Outbound on the server side, private on the client side&lt;/td&gt;
&lt;td&gt;An agent reaches an internal MCP server with neither end on the public internet&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Each one alone leaves a gap. A tunnel terminating at a public gateway keeps the server hidden, but the gateway in front of it is still on the internet. A private gateway on one network doesn't help reach servers in a different network it has no route to, which is the problem tunnels solve. Speakeasy's &lt;a href="https://www.speakeasy.com/blog/release-private-mcp-gateway-tailscale" rel="noopener noreferrer"&gt;private MCP gateway release&lt;/a&gt; describes running both together so that neither end touches the public internet.&lt;/p&gt;

&lt;h2&gt;
  
  
  What are the risks of a public MCP gateway?
&lt;/h2&gt;

&lt;p&gt;A public MCP gateway is not insecure by default. The risks below come from what a public listener makes possible, and from how much then depends on every other control being right:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;It can be found.&lt;/strong&gt; Anything that answers MCP requests on a public address shows up in internet-wide scans, as the Knostic study showed. A gateway hostname also tells an attacker where the organization's tools are concentrated.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pre-authentication code is reachable by everyone.&lt;/strong&gt; TLS termination, OAuth endpoints, discovery metadata, and request parsing all run before a token is validated. A bug in any of them can be exploited from anywhere on the internet.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A stolen token works from any network.&lt;/strong&gt; OAuth access tokens are bearer credentials. If one leaks from a laptop, a log file, or a compromised agent, it can be replayed from anywhere until it expires or is revoked. Against a private gateway, the attacker also needs a device the network admits.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;IP allowlists identify addresses, not people.&lt;/strong&gt; An allowlist narrows who can connect, but the listener stays public and a request carries only a source IP. The organization also needs stable egress addresses for every client and a second list to keep in step with its network policy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Misconfiguration has no backstop.&lt;/strong&gt; If the gateway's authentication or allowlist is configured wrong, its MCP servers are exposed directly, because there is no second boundary behind it.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What are the benefits of a private MCP gateway?
&lt;/h2&gt;

&lt;p&gt;The benefits mirror the risks, with one addition: identity from the network itself.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Nothing to enumerate.&lt;/strong&gt; With no public listener, there is no hostname for a scanner to find and no TLS handshake for a stranger to complete.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two independent checks.&lt;/strong&gt; A caller must be admitted by the network and authenticated by the gateway. A stolen token without an admitted device, or an admitted device without a valid token, gets nowhere.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;User and device identity on each request.&lt;/strong&gt; On overlay networks such as Tailscale, each connection carries the user, device, and tags it came from. The gateway can record that next to the tool call, so the audit trail answers "which machine made this call" as well as "which account."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One place for reachability rules.&lt;/strong&gt; Who can reach the gateway lives in the network policy the organization already maintains, such as a tailnet ACL file or VPC security groups, instead of a separate allowlist in the gateway.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No stable egress requirement.&lt;/strong&gt; Laptops and CI runners don't have to leave the network through fixed IP addresses to satisfy an allowlist.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fails closed.&lt;/strong&gt; When public access is blocked independently of the private network, an overlay outage or a restart makes MCP servers unreachable, never exposed.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What does a private MCP gateway cost?
&lt;/h2&gt;

&lt;p&gt;A private listener has real costs, and for some organizations they outweigh the benefits:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Hosted agents can't reach it.&lt;/strong&gt; Agents that run on a vendor's infrastructure, such as remote connectors in ChatGPT or Claude, call MCP servers from the vendor's network, which isn't on yours. Any MCP server those agents need has to keep a public path. This is why implementations that support private access usually let exposure be set per server rather than for the whole gateway.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Every client needs network membership.&lt;/strong&gt; Developer laptops, CI runners, and background agents all have to be enrolled in the private network. For an organization that already runs Tailscale or a corporate VPN everywhere, that's close to free. For one that doesn't, a private MCP gateway means rolling out a network before rolling out a gateway.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The network becomes a dependency.&lt;/strong&gt; Failing closed is a security property and an availability cost. An overlay outage takes agent tool access down with it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Identity providers usually stay public.&lt;/strong&gt; OAuth flows to an external identity provider still need internet access. AWS &lt;a href="https://aws.amazon.com/blogs/machine-learning/secure-ingress-connectivity-to-amazon-bedrock-agentcore-gateway-using-interface-vpc-endpoints/" rel="noopener noreferrer"&gt;notes this explicitly&lt;/a&gt; for private access to AgentCore Gateway. The gateway's own OAuth endpoints also need to be served on the private hostname, or MCP clients can't complete the authorization flow.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Network access is not authorization.&lt;/strong&gt; A device on the private network is not entitled to every tool. Tool permissions still belong in the gateway, and treating reachability as permission recreates the flat-network problem zero trust networking was meant to fix.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;As a rule of thumb, if most of your agent traffic comes from hosted assistants calling public SaaS MCP servers, a public gateway with strong OAuth is the simpler choice. If most of it comes from agents on company devices calling internal systems, a private gateway fits the way the rest of your network already works.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you deploy a private MCP gateway?
&lt;/h2&gt;

&lt;p&gt;There are three common patterns, ordered from least to most you operate yourself:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Join a managed gateway to your private network.&lt;/strong&gt; The vendor runs the gateway and joins it to your overlay network as a node with a private hostname. This is the least operational work, but it only works with the networks the vendor has integrated with. Speakeasy's Tailscale integration is this pattern.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Use a cloud gateway's private endpoint.&lt;/strong&gt; Cloud-provider gateways can expose a private endpoint inside your VPC. Amazon Bedrock AgentCore Gateway supports an interface VPC endpoint through AWS PrivateLink. Clients inside that cloud get a private path; clients elsewhere need a VPN or peering into the VPC.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Self-host behind an internal load balancer.&lt;/strong&gt; Deploy the gateway in your VPC or data center with no public address and reach it over a VPN or overlay. You get full control of the network path and take on running the gateway, its upgrades, and its availability.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Whichever pattern you choose, the rollout follows the same steps:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Sort your callers.&lt;/strong&gt; List which agents run on devices you manage and which run on vendor infrastructure. The second group decides which MCP servers must keep a public path.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Put the gateway on the network.&lt;/strong&gt; Give it a stable private hostname, and serve the MCP endpoints, OAuth endpoints, and any discovery metadata there so clients can complete authorization without leaving the network.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Write the reachability rules.&lt;/strong&gt; Decide which users, groups, and machine tags can reach the gateway, and express that in the network's own policy. On Tailscale, a grant like the one below lets the engineering group and CI runners reach nodes tagged as the gateway on port 443.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Run both paths while clients move.&lt;/strong&gt; Serve each MCP server on the public and private hostnames at once, repoint clients at the private hostname, and watch until traffic on the public path stops.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Close the public path and test from outside.&lt;/strong&gt; From a machine that is not on the private network, the gateway's public hostname should refuse the request or fail to connect. Repeat the test after load balancer, DNS, or ingress changes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep tool permissions in the gateway.&lt;/strong&gt; Network rules decide who can reach the gateway. The gateway still decides which tools each caller may use.
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"grants"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"src"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"group:eng"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"tag:ci"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"dst"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"tag:mcp-gateway"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"ip"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"tcp:443"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Run from a machine that is not on the private network.&lt;/span&gt;
&lt;span class="c"&gt;# Expect a connection failure or a 403, never a 200 or 401.&lt;/span&gt;
curl -sS -o /dev/null -w &lt;span class="s2"&gt;"%{http_code}\n"&lt;/span&gt; https://mcp.example.com/mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A &lt;code&gt;401&lt;/code&gt; from the public hostname means the listener is still public and is only asking for credentials.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which vendors support a private MCP gateway?
&lt;/h2&gt;

&lt;p&gt;The list is short, and the products differ in shape. Each entry below covers only what the vendor documents.&lt;/p&gt;

&lt;h3&gt;
  
  
  Speakeasy
&lt;/h3&gt;

&lt;p&gt;Speakeasy's &lt;a href="https://www.speakeasy.com/product/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt; can run on a customer's Tailscale tailnet at a stable private hostname, through a feature called private network ingress that Speakeasy &lt;a href="https://www.speakeasy.com/blog/release-private-mcp-gateway-tailscale" rel="noopener noreferrer"&gt;announced in September 2026&lt;/a&gt;. An org admin connects the tailnet with a scoped Tailscale OAuth client, and the gateway joins as a node. Each MCP server is then set to public-only, dual, or private-only. In private-only mode, the platform host and the organization's custom domain return a &lt;code&gt;403&lt;/code&gt; on every public route, and that block doesn't depend on the tailnet being up, so a Tailscale outage or misconfiguration leaves servers unreachable rather than exposed.&lt;/p&gt;

&lt;p&gt;A per-organization attestor forwards only the allowlisted MCP and OAuth paths and attaches the Tailscale user, device, and tags to each request, which then appear in the audit log next to the tool that was called. The MCP endpoint, its OAuth endpoints, and its install page are all served at the private hostname. Paired with Speakeasy's &lt;a href="https://www.speakeasy.com/resources/mcp-tunnels" rel="noopener noreferrer"&gt;MCP tunnels&lt;/a&gt; for servers inside the network, an agent can reach a private MCP server without either end on the public internet.&lt;/p&gt;

&lt;p&gt;Private network ingress is rolling out per organization and is enabled on request. Tailscale is the only overlay network supported today; others are on the roadmap. Per-server authorization using Tailscale ACL capability grants is in development and not yet available.&lt;/p&gt;

&lt;h3&gt;
  
  
  Tailscale Aperture
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://tailscale.com/docs/aperture/mcp-server" rel="noopener noreferrer"&gt;Aperture&lt;/a&gt; is Tailscale's proxy for LLM and MCP traffic on a tailnet. It aggregates tools and resources from configured remote MCP servers behind a single &lt;code&gt;/v1/mcp&lt;/code&gt; endpoint and identifies the caller from the tailnet connection, so requests must come from inside the tailnet and need no separate authorization header. Access is deny-by-default: grants in the Aperture configuration decide which users can list and call which tools. For an organization that has standardized on Tailscale and wants MCP tool access expressed in the same grants model as the rest of its network, Aperture is a more direct fit than a separate gateway.&lt;/p&gt;

&lt;h3&gt;
  
  
  Amazon Bedrock AgentCore Gateway
&lt;/h3&gt;

&lt;p&gt;AgentCore Gateway documents private paths on both sides. For inbound traffic, an &lt;a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/vpc-interface-endpoints.html" rel="noopener noreferrer"&gt;interface VPC endpoint through AWS PrivateLink&lt;/a&gt; gives callers inside a VPC a private route to the gateway. For outbound traffic, a &lt;a href="https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-vpc-egress.html" rel="noopener noreferrer"&gt;VPC Lattice private endpoint&lt;/a&gt; on a gateway target reaches self-hosted MCP servers inside a VPC. AWS notes that OAuth token retrieval with external identity providers, and access to tools hosted outside AWS, still need internet connectivity. It's the natural option for agents that already run inside AWS.&lt;/p&gt;

&lt;h3&gt;
  
  
  Mintive (formerly MintMCP)
&lt;/h3&gt;

&lt;p&gt;Mintive documents two ways to reach MCP servers inside a private network: a private network replica deployed in the customer's VPC that connects outbound, and an outbound-only tunnel to a reverse proxy. Both keep internal MCP servers off the public internet, with traffic flowing from the MCP client to Mintive's policy gateway and then through the replica to the internal server. That is the &lt;a href="https://www.speakeasy.com/resources/mcp-tunnels" rel="noopener noreferrer"&gt;MCP tunnel&lt;/a&gt; pattern, covering the server side of the path. The documentation we reviewed doesn't describe putting the policy gateway's own listener on a private network.&lt;/p&gt;

&lt;h3&gt;
  
  
  Building it yourself
&lt;/h3&gt;

&lt;p&gt;Any MCP gateway you can self-host can be made private by deploying it behind an internal load balancer with no public address and reaching it over a VPN or overlay network. This is the third deployment pattern above, with the same trade: full control of the network path, and full responsibility for the gateway.&lt;/p&gt;

&lt;p&gt;If you're working out which of your MCP servers can come off the public internet and which need to stay reachable for hosted agents, we'd be interested to hear how you're drawing that line. &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;Talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A private MCP gateway is a Model Context Protocol (MCP) gateway whose listener is reachable only from a private network, such as a VPN, a Tailscale tailnet, or a cloud VPC. Agents reach it over that network, and the public internet has no path to the MCP servers behind it. The term describes the gateway's exposure, meaning who can open a connection to it, rather than where it runs or who operates it.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is a private MCP gateway different from an MCP gateway?
&lt;/h3&gt;

&lt;p&gt;An MCP gateway is the policy point between AI agents and MCP servers: it authenticates callers, decides which tools each one may use, and logs every tool call. A private MCP gateway does the same job but listens only on a private network. That adds a second check in front of the gateway's own authentication, because a caller has to be admitted by the network before a request can reach the gateway at all.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is the difference between a self-hosted and a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Self-hosted describes who operates the gateway software; private describes who can reach the gateway. The two are independent. A self-hosted MCP gateway behind a public load balancer is exposed to the internet, and a managed MCP gateway run by a vendor can be private if the vendor joins it to the customer's private network. Deploying a self-hosted gateway behind an internal load balancer makes it private, but that is a property of the deployment, not of self-hosting.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is an on-prem MCP gateway the same as a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;No. An on-prem MCP gateway runs on hardware in the organization's own facilities, which is a statement about location. On-prem networks routinely publish services to the internet, so an on-prem gateway is private only if its listener is configured to be reachable from the internal network alone. Many on-prem gateways are private in practice, but the posture still has to be set and verified.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is a private MCP gateway the same as an MCP tunnel?
&lt;/h3&gt;

&lt;p&gt;No. An MCP tunnel is an outbound-only connection that lets a gateway or agent platform reach an MCP server inside a private network without opening an inbound port, so it makes a private server reachable. A private MCP gateway keeps the gateway's own listener off the public internet, so the path from agents to the gateway is private. The two compose: with a tunnel on the server side and a private gateway on the client side, an agent can reach an internal MCP server without either end touching the public internet.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is putting an MCP gateway behind a VPN the same as a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A VPN is one way to build a private MCP gateway, along with overlay networks such as Tailscale and private cloud endpoints such as AWS PrivateLink. The gateway only counts as private if the VPN or overlay is the only way to reach it. A gateway that is reachable over a VPN and also still answers on a public hostname has a private path, but it is not a private MCP gateway.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is an IP allowlist enough to make an MCP gateway private?
&lt;/h3&gt;

&lt;p&gt;No. An IP allowlist narrows which addresses can connect to a public MCP gateway, but the listener stays on the public internet, so the gateway's TLS stack and OAuth endpoints are still reachable by anyone who can scan for them. A request through an allowlist carries only a source IP, and the organization needs stable egress addresses and a second list to keep in step with its network policy. A private MCP gateway removes the public listener instead of filtering it.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are the risks of a public MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A public MCP gateway can be found by internet-wide scans, exposes its pre-authentication code such as TLS termination and OAuth endpoints to everyone, and accepts a stolen access token from any network until the token expires or is revoked. If its authentication or allowlist is misconfigured, there is no second boundary behind it. None of these make a well-run public gateway insecure, but each one depends on every other control being right.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are the benefits of a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;A private MCP gateway has no public listener to enumerate, requires every caller to pass the private network's admission and the gateway's authentication, and moves reachability policy into the network rules an organization already maintains. On overlay networks such as Tailscale, each request can also carry the user and device it came from, which the gateway can record next to the tool call. A well-built private gateway fails closed, so an outage makes MCP servers unreachable rather than exposed.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are the downsides of a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Agents that run on a vendor's infrastructure, such as remote connectors in ChatGPT or Claude, cannot reach a gateway that listens only on a private network. Every other client, including developer laptops, CI runners, and background agents, has to be enrolled in that network. The network also becomes a dependency, so an overlay outage takes agent tool access with it, and OAuth flows to an external identity provider usually still need internet access.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can ChatGPT or Claude connect to a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Clients that call MCP servers from the vendor's cloud, such as remote connectors in ChatGPT and Claude, cannot reach a gateway that listens only on a private network, because the vendor's infrastructure is not on that network. Clients that run on a device inside the network, such as Claude Code or Cursor on an enrolled laptop, can. Organizations that need both usually keep a public path for the specific MCP servers hosted agents use and make the rest private-only.&lt;/p&gt;

&lt;h3&gt;
  
  
  Does a private MCP gateway replace MCP authentication?
&lt;/h3&gt;

&lt;p&gt;No. Network access decides who can reach the gateway; MCP authentication and tool permissions decide what a connected caller may do. A device on the private network is not entitled to every tool, so a private MCP gateway still authenticates each MCP client with OAuth and enforces per-tool policy. Treating network reachability as authorization recreates the flat-network problem that zero trust networking was meant to fix.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do you deploy a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;There are three common patterns: join a managed MCP gateway to your private network as a node, use a cloud gateway's private endpoint inside your VPC, or self-host a gateway behind an internal load balancer and reach it over a VPN or overlay network. In each case, serve the MCP and OAuth endpoints on a private hostname, write network rules for who can reach it, run public and private paths side by side while clients move, then close the public path and test from outside the network that it refuses requests.&lt;/p&gt;

&lt;h3&gt;
  
  
  Which vendors support a private MCP gateway?
&lt;/h3&gt;

&lt;p&gt;Speakeasy can run its MCP gateway on a customer's Tailscale tailnet at a private hostname, with each MCP server set to public-only, dual, or private-only; the feature is rolling out per organization. Tailscale Aperture proxies MCP servers behind a single endpoint reached from inside the tailnet. Amazon Bedrock AgentCore Gateway supports a PrivateLink interface endpoint for inbound traffic and VPC Lattice private endpoints for MCP targets. Mintive documents private-network replicas and tunnels for reaching MCP servers inside a VPC, which is closer to the MCP tunnel pattern.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>infrastructure</category>
      <category>mcp</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What is an AI security server?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Wed, 11 Nov 2026 10:52:36 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-an-ai-security-server-3180</link>
      <guid>https://dev.to/speakeasy/what-is-an-ai-security-server-3180</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/ai-security-server" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  AI security server
&lt;/h2&gt;

&lt;p&gt;An AI security server is the HTTPS service that a model provider calls, with the transcript of a governed request, before inference runs, so the organization can allow or deny that request under its own policy. It runs on the provider's request path rather than on the user's device or network, which is why one endpoint governs every user on every covered surface.&lt;/p&gt;

&lt;p&gt;The term is &lt;a href="https://www.anthropic.com" rel="noopener noreferrer"&gt;Anthropic's&lt;/a&gt;. &lt;a href="https://platform.claude.com/docs/en/manage-claude/inference-hooks" rel="noopener noreferrer"&gt;Inference hooks&lt;/a&gt;, a Claude Enterprise feature in beta at the time of writing, hold each governed prompt, POST the conversation to an endpoint the organization names, and proceed only on &lt;code&gt;allow&lt;/code&gt;. Anthropic's documentation calls that endpoint the AI security server, whether the customer runs it or a security vendor runs it on the customer's behalf. Inference hooks are the only implementation of the pattern we know of from a major model provider, so this page describes the pattern with Anthropic's version as the worked example, and the details below should be re-checked against Anthropic's docs before anyone builds on them.&lt;/p&gt;

&lt;p&gt;The pattern exists because the controls a security team already operates stop at the edge of the hosted model. Email DLP reads outbound mail, web DLP inspects uploads through the forward proxy, and endpoint DLP watches file copies and clipboard events. A prompt typed into claude.ai crosses none of those in a form they can read; from the network's point of view it is a TLS session from a sanctioned application to a sanctioned domain. The &lt;a href="https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-agent-governance-framework-gap-20260403/" rel="noopener noreferrer"&gt;Cloud Security Alliance&lt;/a&gt; put numbers on the resulting gap in its April 2026 research note: in the survey it cites, 71% of 235 large-enterprise CISOs and CIOs said AI already has access to core business systems, and 16% said that access is governed effectively.&lt;/p&gt;

&lt;p&gt;An AI security server has five parts. The first three define the protocol, and the last two are what let an organization turn it on without blocking anyone on day one:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;A signed request.&lt;/strong&gt; The provider sends an HTTPS &lt;code&gt;POST&lt;/code&gt; for each governed inference request. Anthropic signs it per the &lt;a href="https://www.standardwebhooks.com" rel="noopener noreferrer"&gt;Standard Webhooks&lt;/a&gt; specification with a secret the organization generates once, and can attach up to 16 static custom headers so the server can authenticate the caller.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The transcript, as the user sees it.&lt;/strong&gt; The body carries the conversation up to the point of inference: text, tool calls and their results, text extracted from attachments, and prior turns, plus the actor, the originating application, and the model.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A verdict.&lt;/strong&gt; The server returns HTTP 200 with a small JSON body. &lt;code&gt;{"action": "allow"}&lt;/code&gt; lets inference proceed. A deny carries a &lt;code&gt;deny_reason&lt;/code&gt; the end user sees and an optional &lt;code&gt;reference_id&lt;/code&gt; the organization can join to its own records.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A timeout and a failure posture.&lt;/strong&gt; The provider waits a configured interval for the verdict. If none arrives, the organization's failure handling setting decides whether the request is blocked or proceeds uninspected.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rollout controls.&lt;/strong&gt; Shadow mode returns verdicts without blocking anything, a rollout percentage inspects a fraction of requests, and role exclusions exempt chosen groups, so enforcement can start at zero and climb.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  How does an AI security server work?
&lt;/h2&gt;

&lt;p&gt;The round trip is synchronous and sits in the user's critical path. A user submits a prompt on a governed surface. Before inference begins, the provider holds the request and POSTs the transcript to the configured endpoint. The server evaluates it against whatever policy the organization runs, which might be a regex for API keys, a PII classifier, a model allowlist, a working-hours rule, or an LLM judge, and answers. On &lt;code&gt;allow&lt;/code&gt;, inference proceeds. On &lt;code&gt;deny&lt;/code&gt;, the request is rejected before the model sees it, and the user reads a blocked-by-policy message assembled from the server's &lt;code&gt;deny_reason&lt;/code&gt; followed by a standing message the administrator configured, such as who to contact for an exception.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"action"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"deny"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"deny_reason"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"This prompt appears to contain customer payment card data, which your organization's policy does not allow."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"reference_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"scan_01HXPT4R9V"&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Availability engineering is most of the operational work. On Claude Enterprise the verdict timeout is configurable from 1 to 10,000ms with a 5,000ms default, and the budget covers the whole exchange, including connection and TLS handshake. A slow verdict counts as an unreachable server. Anthropic retries once, after 100ms, and only when the connection attempt itself failed; once the server has responded, the exchange is never retried. Sustained failures trip a circuit breaker: Anthropic stops contacting the server and applies the failure posture to every request, then checks in the background, starting ten minutes after the trip, for a valid verdict before resuming. A non-200 status, a redirect, or an unparseable body is a webhook failure rather than a deny, so a server cannot signal "block this" with an error code.&lt;/p&gt;

&lt;p&gt;Because the check runs on the provider's infrastructure, there is nothing to install on user devices, no proxy to insert in front of claude.ai, and no client version to keep current. One administrator setting applies to every user on every governed surface the same day. That is the property that makes the pattern attractive to security teams, and it is also the source of every limit described below, because the server sees exactly what the provider chooses to send and nothing else.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why not point the hook at the DLP scanner you already run?
&lt;/h2&gt;

&lt;p&gt;For many organizations that is the right first move, and Anthropic says DLP forwarding is the most common deployment it sees. The scanner already knows the organization's data classes. Its regexes for card numbers, national identifiers, and API-key formats work as well on a chat transcript as on an email body. The security team already trusts its verdicts and already has a workflow for its findings. If the whole policy is "no regulated data enters a hosted model," a DLP engine behind an HTTPS endpoint answers that question in an afternoon of integration work, and a fifty-line always-allow server that archives every transcript is a legitimate deployment on its own, as a push-based alternative to polling the Compliance API.&lt;/p&gt;

&lt;p&gt;The scanner starts to strain when the policy needs more than a data class. An engine built for mail returns a verdict per message with no memory of the conversation, so it cannot distinguish a user pasting a customer record into a fresh chat from a tool result that returned one three turns ago. It has no concept of which connector produced a tool result, so it cannot say that a payload from the Salesforce tool is fine while the same payload from an unknown MCP server is not. Its findings land in a DLP console, separate from the transcript that produced them, which means an investigator joins two systems to reconstruct one event. And the transcript arrives untruncated, so a scanner behind a proxy with a 1 MB body limit will reject long conversations, and under fail-open handling those rejected prompts reach the model uninspected.&lt;/p&gt;

&lt;p&gt;The scanner is still worth deploying. The useful exercise is naming which question each endpoint can answer. A DLP engine answers "did a forbidden data class appear." A purpose-built server, or an &lt;a href="https://www.speakeasy.com/resources/ai-control-plane" rel="noopener noreferrer"&gt;AI Control Plane&lt;/a&gt; acting as the endpoint, answers "given who this user is, what tool produced this content, and what the rest of the conversation says, should this request proceed," and it can apply the same policy engine to the tool and MCP traffic the hook never sees. &lt;a href="https://www.speakeasy.com/resources/agent-dlp" rel="noopener noreferrer"&gt;Agent DLP&lt;/a&gt; is the broader control both are enforcing.&lt;/p&gt;

&lt;h2&gt;
  
  
  AI security server vs classic DLP vs AI agent hooks: what's the difference?
&lt;/h2&gt;

&lt;p&gt;Three controls that all claim to govern AI usage sit in three different places, and the place determines what each can see.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Control&lt;/th&gt;
&lt;th&gt;Where it runs&lt;/th&gt;
&lt;th&gt;What it sees&lt;/th&gt;
&lt;th&gt;What it governs&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Classic DLP / CASB / SSE&lt;/td&gt;
&lt;td&gt;Mail flow, forward proxy, endpoint agent&lt;/td&gt;
&lt;td&gt;Files, email, web uploads, endpoint actions&lt;/td&gt;
&lt;td&gt;The channels it proxies&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI agent hooks&lt;/td&gt;
&lt;td&gt;Inside the agent, on the user's machine&lt;/td&gt;
&lt;td&gt;Prompts, tool calls before they run, tool results&lt;/td&gt;
&lt;td&gt;One agent, per installation&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;AI security server&lt;/td&gt;
&lt;td&gt;The provider's request path&lt;/td&gt;
&lt;td&gt;The conversation transcript, before inference&lt;/td&gt;
&lt;td&gt;Every governed user, from one setting&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Classic DLP inspects channels the organization controls; an AI security server inspects a request the provider controls. The first sees a prompt as encrypted traffic to an approved domain. The second receives it decoded, from the party that was about to act on it. When an organization forwards the hook to its DLP engine, the engine's detection logic is reused, but the vantage point is new.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/ai-agent-hooks" rel="noopener noreferrer"&gt;AI agent hooks&lt;/a&gt; are the closer comparison, because the vocabulary overlaps and both fire on prompts. A hook is a handler that runs inside the agent, on the machine that hosts it, at lifecycle events such as &lt;code&gt;UserPromptSubmit&lt;/code&gt;, &lt;code&gt;PreToolUse&lt;/code&gt;, and &lt;code&gt;PostToolUse&lt;/code&gt;. Claude Code, Cursor, Codex, and VS Code Copilot all expose them. A hook sees a shell command before it executes and can block it; it governs the one agent it was configured for; and it has to be present on every machine, because an agent launched without the configuration is ungoverned. An AI security server trades depth for breadth. It sees inference requests rather than tool invocations, so it cannot stop &lt;code&gt;rm -rf&lt;/code&gt; before it runs, but it covers every user in the organization from a single admin setting, on surfaces such as claude.ai where no client-side hook exists.&lt;/p&gt;

&lt;p&gt;The two compose. Hooks govern coding agents on developer machines, where the consequential actions are local. The server governs the chat and web surfaces that make up most of a non-engineering workforce's Claude usage. Anthropic's Compliance API is the after-the-fact complement to both; it retrieves what happened, and an AI security server decides whether it happens.&lt;/p&gt;

&lt;h2&gt;
  
  
  What are Anthropic Inference hooks?
&lt;/h2&gt;

&lt;p&gt;Inference hooks are the Claude Enterprise feature that calls an AI security server. Anthropic labels the feature beta and warns that field names, request shapes, and headers may change, so everything in this section is a snapshot of the docs at the time of writing rather than a contract.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;One hook, many surfaces.&lt;/strong&gt; A single organization-level endpoint governs conversations across claude.ai, Cowork, and Claude Code, whether they run on the web, in the desktop or mobile apps, or in the CLI.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Prompt side only.&lt;/strong&gt; The only event is &lt;code&gt;prompt&lt;/code&gt;, which fires once per governed inference request. Anthropic describes response-side enforcement as a planned later event.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Allow or deny.&lt;/strong&gt; Verdicts cannot rewrite or redact. A prompt carrying a secret is blocked whole or passed whole.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Owner-level permission.&lt;/strong&gt; Configuration requires &lt;code&gt;organization:manage&lt;/code&gt;, held only by the Owner and Primary owner roles. The Admin role is not sufficient.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Three enforcement states.&lt;/strong&gt; Off, shadow, and enforcing. Allowing the feature for the organization always forces enforcement off first, so turning it on never starts inspection by itself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rollout controls.&lt;/strong&gt; A rollout percentage from 0 to 100, rolled once per conversation turn; exclusions for custom roles the organization created (built-in roles cannot be excluded, and machine-credential traffic is always inspected); and a custom blocked-prompt message of up to 500 characters.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Out of scope.&lt;/strong&gt; Amazon Bedrock, Google Cloud, Platform organizations using the Claude API, and voice mode are not covered. Ancillary requests such as title generation are never sent.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The endpoint requirements are firm. The URL must be &lt;code&gt;https://&lt;/code&gt; on port 443, on a publicly routable host, with a certificate that validates against the public CA store, reachable without redirects. Reverse-tunnel services such as ngrok are blocked by Anthropic's network policy, so testing happens on a domain the organization controls. Verdicts must return as HTTP 200. Requests originate from a published Anthropic egress range, which can be allowlisted, though Anthropic notes that allowlisting does not replace signature verification.&lt;/p&gt;

&lt;p&gt;Anthropic's &lt;a href="https://platform.claude.com/docs/en/manage-claude/inference-hooks-configuration" rel="noopener noreferrer"&gt;Configure Inference hooks&lt;/a&gt; and &lt;a href="https://platform.claude.com/docs/en/manage-claude/inference-hooks-endpoint" rel="noopener noreferrer"&gt;Develop an Inference hooks integration&lt;/a&gt; pages carry the request and verdict schemas, the signature procedure, and the operational semantics, and they are the source to trust over this page when the two disagree.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does an AI security server not see?
&lt;/h2&gt;

&lt;p&gt;The server's view is bounded by what the provider sends, and the gaps shape the policy more than the contents do. On Claude Enterprise today:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Tool calls before they run.&lt;/strong&gt; The transcript carries &lt;code&gt;tool_use&lt;/code&gt; blocks and their results, so the server reads them, but the hook gates inference requests rather than tool invocations. A destructive connector call is visible in the turn that reports its result, after it has already executed. The server can deny the next inference on the strength of that result; it cannot prevent the call.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tool definitions and system prompts.&lt;/strong&gt; Neither is ever included. A &lt;a href="https://www.speakeasy.com/resources/mcp-tool-poisoning" rel="noopener noreferrer"&gt;poisoned tool description&lt;/a&gt; that steers the model is invisible to this control, which is one reason tool-path governance stays a separate problem.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Attachments as bytes.&lt;/strong&gt; Attachments arrive as metadata and extracted text. Raw file and image bytes are never sent, so image-only content, such as a screenshot of a spreadsheet, is not inspected.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude's hidden reasoning and Anthropic-internal context.&lt;/strong&gt; Only the conversation as the user sees it is sent.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sampled-out and excluded traffic.&lt;/strong&gt; A request outside the rollout percentage or from a member of an excluded role is never sent, so the server can neither evaluate nor record it. Under fail-closed handling, sampled-out requests still proceed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Everything off the governed surfaces.&lt;/strong&gt; Another provider's model, a personal Claude account, a Bedrock deployment, or a local agent calling an MCP server directly all sit outside this path. That is the &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;shadow AI&lt;/a&gt; problem, and an AI security server does not solve it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are Anthropic's stated limits at the time of writing, and the beta is still moving. Check the &lt;a href="https://platform.claude.com/docs/en/manage-claude/inference-hooks" rel="noopener noreferrer"&gt;Inference hooks overview&lt;/a&gt; before treating any of them as permanent.&lt;/p&gt;

&lt;h2&gt;
  
  
  What are the trade-offs?
&lt;/h2&gt;

&lt;p&gt;An AI security server adds its round trip to the latency of every governed request in the organization. A 200ms verdict on a 3-second generation is invisible; a 4-second verdict on a one-line question is not. Anthropic advises load-testing the server before a large rollout, and a server that calls an LLM judge for every prompt should expect to spend most of its timeout budget doing so.&lt;/p&gt;

&lt;p&gt;Fail-closed enforcement couples Claude's availability to the server's uptime. If the endpoint is down and failure handling is set to block, every governed user in the organization is blocked until the circuit breaker resets or an administrator intervenes. Fail-open removes that coupling and reintroduces the gap the server was meant to close, since every request during an outage reaches the model uninspected. There is no setting that avoids both costs; the choice is which failure the organization prefers.&lt;/p&gt;

&lt;p&gt;The protocol is one vendor's beta. Anthropic reserves the right to change field names and request shapes, the verdict vocabulary is two words, and no other major provider has shipped an equivalent. An organization that builds policy logic against this contract should expect to revise it, and should not assume the same server will govern a second provider's traffic without a second integration. Allow-or-deny also means a blocked prompt is blocked whole, so a user who pasted one card number into a long, otherwise legitimate request loses the whole request and has to try again.&lt;/p&gt;

&lt;h2&gt;
  
  
  Do I need an AI security server?
&lt;/h2&gt;

&lt;p&gt;You need one if three things are true at once: your organization runs Claude Enterprise, it has a written policy about what may enter a hosted model, and no control you operate today can enforce that policy on claude.ai, Cowork, or Claude Code on the web. The third is true by default on the web surfaces, because the existing DLP stack cannot read the traffic and there is no client to install a hook into, so the decision usually turns on whether the second exists in writing.&lt;/p&gt;

&lt;p&gt;You do not need one if audit is the requirement rather than enforcement. The Compliance API retrieves transcripts and activity after the fact, and a records-retention or investigations workflow is well served by it. You also do not need one for coding agents on managed developer machines if agent hooks are already deployed there, though the server still covers the surfaces hooks cannot reach.&lt;/p&gt;

&lt;p&gt;If the answer is yes, the decision that follows is which endpoint. Choose the DLP engine you already run if the policy is a set of data classes and the findings workflow already exists. Choose a purpose-built server or an AI Control Plane if the policy depends on conversation context, tool provenance, or user identity, if findings should sit on the transcript that produced them, or if the same policy engine has to govern the MCP and tool path as well.&lt;/p&gt;

&lt;h2&gt;
  
  
  A note on Speakeasy
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://www.speakeasy.com/product/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt; can be configured as the AI security server for a Claude Enterprise organization. The &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/org-admin/ai-integrations/anthropic-inference-hooks" rel="noopener noreferrer"&gt;Anthropic inference hooks&lt;/a&gt; docs and the &lt;a href="https://www.speakeasy.com/use-cases/claude-control" rel="noopener noreferrer"&gt;Claude security use case&lt;/a&gt; walk through the setup; in outline, Speakeasy generates an organization-specific HTTPS endpoint, an Owner pastes it into Claude's Inference hooks settings, and the signing secret Claude reveals once is saved back into Speakeasy.&lt;/p&gt;

&lt;p&gt;Each governed prompt is evaluated against the project's &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/secure/guardrails" rel="noopener noreferrer"&gt;risk policies&lt;/a&gt;, using the same &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/secure/detection-rules" rel="noopener noreferrer"&gt;detection rules&lt;/a&gt; and classifiers that cover the organization's other agent traffic, and a denied request never reaches the model. Every turn lands in &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/observe/agent-sessions" rel="noopener noreferrer"&gt;Agent Sessions&lt;/a&gt; with the verdict attached and the person behind the prompt resolved through the identity provider, so an investigator reads one record rather than joining a findings log to a chat export. Tool calls, connectors, and MCP servers are governed separately through the &lt;a href="https://www.speakeasy.com/product/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt;, which terminates each call with the caller's identity attached and can block or redact before forwarding. That is the surface Inference hooks do not gate, and it is where the destructive actions live.&lt;/p&gt;

&lt;p&gt;Two limits apply. Speakeasy is one of several possible endpoints, and a DLP engine the organization already trusts is a reasonable alternative for a data-class-only policy. The inference hook is also the broadest control Speakeasy offers for Claude rather than the deepest; for Claude Code and Cowork, the &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/org-admin/logging-and-telemetry" rel="noopener noreferrer"&gt;logging and telemetry&lt;/a&gt; integration and &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/plugins/anthropic" rel="noopener noreferrer"&gt;managed Anthropic plugins&lt;/a&gt; add tool-level visibility and hooks that see a tool call before it runs. The recommended rollout mirrors Anthropic's own safeguards: shadow mode at 100% inspection, tune policies against live verdicts, then switch failure handling to block.&lt;/p&gt;

&lt;p&gt;If you want to see what your Claude Enterprise traffic would have produced under these policies last week, &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is an AI security server?
&lt;/h3&gt;

&lt;p&gt;An AI security server is the HTTPS service that a model provider calls, with the transcript of a governed request, before inference runs, so the organization can allow or deny that request under its own policy. It runs on the provider's request path rather than on the user's device or network, which is why one endpoint governs every user on every covered surface. The term comes from Anthropic's Inference hooks documentation for Claude Enterprise, where it names the endpoint the customer or its security vendor operates.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does an AI security server work?
&lt;/h3&gt;

&lt;p&gt;A user submits a prompt on a governed surface. The provider holds the request and sends a signed HTTPS POST carrying the conversation transcript to the endpoint the organization configured. The server evaluates the transcript against the organization's policy and returns a JSON verdict, allow or deny, within a configured timeout. On allow the request proceeds to the model. On deny the user sees a blocked-by-policy message and the model never receives the prompt.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is an AI security server different from classic DLP?
&lt;/h3&gt;

&lt;p&gt;Classic DLP, CASB, and SSE products inspect the channels they proxy: outbound email, web uploads, file transfers, and endpoint actions. A prompt sent to a hosted model crosses those channels as a TLS session from a sanctioned application to a sanctioned domain, so the content is never readable there. An AI security server receives the decoded transcript from the provider itself, before the model runs. Many organizations point the hook at a DLP engine they already own, so the two are often the same product playing a new role rather than competitors.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is an AI security server different from AI agent hooks?
&lt;/h3&gt;

&lt;p&gt;AI agent hooks run inside the agent on the machine where it executes, so a hook in Claude Code, Cursor, or Codex sees a shell command before it runs and governs the one agent it was configured for. An AI security server runs on the provider's side of the wire, sees inference requests rather than tool invocations, and governs every user in the organization from one admin setting. The two compose: hooks cover coding agents on developer machines in depth, and the server covers chat and web surfaces where no client-side control exists.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are Anthropic Inference hooks?
&lt;/h3&gt;

&lt;p&gt;Inference hooks are the Claude Enterprise feature that routes every governed prompt through an AI security server before inference. At the time of writing the feature is in beta. One organization-level endpoint covers claude.ai, Cowork, and Claude Code sessions on the web, in the desktop and mobile apps, and in the CLI. The only event is prompt, which fires once per governed inference request, and verdicts are allow or deny with no rewrite or redaction. Anthropic's Inference hooks documentation is the authoritative source for current scope.&lt;/p&gt;

&lt;h3&gt;
  
  
  What does an AI security server not see?
&lt;/h3&gt;

&lt;p&gt;On Claude Enterprise the server receives the transcript as the user sees it: text, tool calls and their results, and text extracted from attachments. It never receives raw file or image bytes, system prompts, tool definitions, or Anthropic-internal context. It cannot stop a tool call before it runs, because the hook gates inference requests rather than tool invocations. Requests outside the rollout percentage, from excluded roles, on Amazon Bedrock or Google Cloud, through the Claude Platform API, or in voice mode are never sent.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can an AI security server redact or rewrite a prompt?
&lt;/h3&gt;

&lt;p&gt;Not on Claude Enterprise today. Anthropic's Inference hooks support two verdicts, allow and deny, so a prompt that contains a secret is blocked whole or passed through whole. Redaction is a property of controls that sit in the data path and can modify the payload, such as an MCP gateway inspecting a tool result. Anthropic lists rewriting and redaction as current limitations rather than design goals, so this may change, but nothing should be planned around it yet.&lt;/p&gt;

&lt;h3&gt;
  
  
  What happens if the AI security server goes down?
&lt;/h3&gt;

&lt;p&gt;The organization's failure handling setting decides. Fail closed blocks every governed request until the server returns; fail open lets requests proceed to the model without inspection. Sustained failures trip a circuit breaker, after which Anthropic stops contacting the server and applies the failure posture to every request until the breaker resets, either when an administrator turns enforcement back on or when a background health check gets a valid verdict. Fail-closed mode therefore couples Claude's availability to the server's uptime.&lt;/p&gt;

&lt;h3&gt;
  
  
  Should I point Inference hooks at my existing DLP scanner or run a purpose-built server?
&lt;/h3&gt;

&lt;p&gt;If the only question is whether a regulated data class entered a prompt, the scanner already running on email is a reasonable endpoint, and Anthropic reports DLP forwarding as the most common deployment. A purpose-built server or an AI Control Plane earns its place when the policy needs conversation context, when findings should sit next to the transcript that produced them, or when the same policy engine also has to govern the tool and MCP path the hook does not reach.&lt;/p&gt;

&lt;h3&gt;
  
  
  Do I need an AI security server?
&lt;/h3&gt;

&lt;p&gt;You need one if your organization runs Claude Enterprise, has a policy about what may enter a hosted model, and has no control today that can enforce that policy on claude.ai, Cowork, or Claude Code on the web. Reading the Compliance API after the fact is enough for audit. If you have to stop a request rather than record it, an AI security server is the only place on those surfaces where that decision can be made.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is an AI security server the same thing as an AI gateway?
&lt;/h3&gt;

&lt;p&gt;No. An AI gateway is a proxy the organization runs between its own applications and model APIs, so traffic must be routed through it deliberately. An AI security server is called by the provider, so it covers the provider's first-party apps, which never pass through a gateway the customer controls. A gateway can be the AI security server's endpoint if it exposes the verdict protocol, but the two govern different traffic.&lt;/p&gt;

&lt;h3&gt;
  
  
  Where does Speakeasy fit?
&lt;/h3&gt;

&lt;p&gt;The Speakeasy AI Control Plane can be configured as the AI security server for a Claude Enterprise organization. Each governed prompt is evaluated against the project's risk policies and detection rules, a denied request never reaches the model, and every turn is recorded in Agent Sessions with the verdict attached and the user resolved through the identity provider. The platform also governs the tool and MCP path through its MCP gateway, which is the surface Inference hooks do not gate. It is one of several possible endpoints, not the only one.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>security</category>
    </item>
    <item>
      <title>What is an agent plugin?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Wed, 11 Nov 2026 10:03:13 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-is-an-agent-plugin-2hl1</link>
      <guid>https://dev.to/speakeasy/what-is-an-agent-plugin-2hl1</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/agent-plugin" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Agent plugin
&lt;/h2&gt;

&lt;p&gt;An agent plugin is an installable package that extends an AI agent by bundling&lt;br&gt;
reusable capabilities: typically skills (workflow instructions), one or more&lt;br&gt;
Model Context Protocol (MCP) servers or connectors, and harness-specific extras&lt;br&gt;
such as hooks, rules, commands, or subagents, so a team can distribute a whole&lt;br&gt;
capability in one install rather than wiring each piece by hand.&lt;/p&gt;

&lt;p&gt;Through the Agent Plugins standard, an MCP host exposes capabilities. Cursor, Claude, and ChatGPT use a plugin to bundle that backend and related abilities, so any admin or user can grab it from a repo or marketplace.&lt;/p&gt;

&lt;p&gt;What makes it a complication: "plugin" isn't a single thing. Each harness, meaning the application that runs the agent and decides what loads, has its own manifest, its own component list, and its own marketplace, and the same word covers meaningfully different things across &lt;a href="https://cursor.com/docs/plugins" rel="noopener noreferrer"&gt;Cursor&lt;/a&gt;, &lt;a href="https://docs.claude.com/en/docs/claude-code/plugins" rel="noopener noreferrer"&gt;Claude&lt;/a&gt;, and &lt;a href="https://developers.openai.com/plugins/concepts/plugins" rel="noopener noreferrer"&gt;ChatGPT&lt;/a&gt;. A &lt;a href="https://agent-plugins.org" rel="noopener noreferrer"&gt;Agent Plugins&lt;/a&gt; spec defines one portable base all 3 now use more, while every harness still adds its own extras. This entry defines the shared idea, then spells out how each harness uses it.&lt;/p&gt;

&lt;h2&gt;
  
  
  How is an agent plugin different from an MCP server?
&lt;/h2&gt;

&lt;p&gt;A &lt;a href="https://modelcontextprotocol.io" rel="noopener noreferrer"&gt;server for MCP&lt;/a&gt; is software that’s active. It exposes functions, data, and prompts to a program through a protocol, runs on a channel (STDIO locally or Streamable HTTP at a URL), and performs the real task when the program invokes a function. A plugin isn’t an active application. It’s a directory holding its manifest, and contains an &lt;code&gt;mcp.json&lt;/code&gt; entry that tells the system which hosts to use and how.&lt;/p&gt;

&lt;p&gt;MCP shows what can be done and how. Plugins solve distribution: how those tools, plus the &lt;a href="https://www.speakeasy.com/blog/skills-vs-mcp" rel="noopener noreferrer"&gt;skills&lt;/a&gt; that teach the agent to use them well, plus whatever harness-specific configuration the workflow needs, reach a hundred machines with one install and stay updated together under one version.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is the Agent Plugins standard?
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://agent-plugins.org" rel="noopener noreferrer"&gt;Agent Plugins&lt;/a&gt; specification is vendor-neutral and covers the software, issued at 1.0 during August 2026 from a team with maintainers at Amazon, Cursor, Microsoft, and OpenAI, plus Vercel. It fixes a setup any compatible harness can use: up top, a &lt;code&gt;plugin.json&lt;/code&gt; manifest that identifies the plugin and pins its spec, plus an &lt;code&gt;mcp.json&lt;/code&gt; with definitions for Streamable HTTP transports and STDIO, and folders in the &lt;code&gt;skills/&lt;/code&gt; directory using &lt;a href="https://agentskills.io" rel="noopener noreferrer"&gt;Agent Skills&lt;/a&gt; style. Harness-specific actions sit above it under reverse-domain namespaces like &lt;code&gt;com.example.client&lt;/code&gt;, keeping the portable part portable.&lt;/p&gt;

&lt;p&gt;The spec is kept tight. It standardizes the bundle, while each harness manages setup, sharing, marketplaces, permissions, and credential use. That divide explains why the parts below still vary. One conformant add-on installs for Cursor, Codex, ChatGPT, GitHub Copilot as well as Kiro and VS Code as is, but this harness sets which parts run later, who checked its page, and if any admin can make it mandatory. Claude plugins have a separate manifest, detailed below.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is a Cursor plugin?
&lt;/h2&gt;

&lt;p&gt;As the &lt;a href="https://cursor.com/docs/plugins" rel="noopener noreferrer"&gt;docs for Cursor plugins&lt;/a&gt; note, both formats work in Cursor:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Agent Plugins&lt;/strong&gt; are spec-conformant packages in the format described above, with &lt;code&gt;plugin.json&lt;/code&gt; at the plugin root, carrying skills and MCP servers. Cursor has read the portable format since version 3.13, so any conformant plugin loads without changes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cursor Plugins&lt;/strong&gt; use a &lt;code&gt;.cursor-plugin/plugin.json&lt;/code&gt; manifest and add Cursor-specific components: rules (&lt;code&gt;.mdc&lt;/code&gt; guidance files), custom agents, commands, hooks, and variables.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Both come through Cursor Marketplace, with every page checked before it goes live, or via internal marketplaces for Teams and Enterprise. Marketplaces handle control: admins limit who gets every plugin and pick Required or either Default Off or Default On; under Required, it stays in place and can't be uninstalled.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is a Claude plugin?
&lt;/h2&gt;

&lt;p&gt;A Claude plugin is a directory identified by a &lt;code&gt;.claude-plugin/plugin.json&lt;/code&gt; manifest that bundles skills, connectors (Anthropic's word for MCP servers in its apps), subagents, and hooks. In &lt;a href="https://docs.claude.com/en/docs/claude-code/plugins" rel="noopener noreferrer"&gt;Claude Code&lt;/a&gt;, the same package can also carry LSP servers, background monitors, executables, and default settings, and plugins install from GitHub-backed marketplaces: Anthropic's official and community catalogs, or a private repository a team controls.&lt;/p&gt;

&lt;p&gt;In the Claude apps, the &lt;a href="https://support.claude.com/en/articles/13837440-use-plugins-in-claude" rel="noopener noreferrer"&gt;plugin surface&lt;/a&gt; is thinner in chat than in Claude Cowork. A plugin installs from the Customize menu, and its skills work in chat on the web, the desktop Chat tab, and Cowork, while hooks and subagents run only in Cowork (and in Claude Code). With Team and Enterprise, admins distribute plugins via company marketplaces and can set one as mandatory, so it installs for all users and prevents uninstalls.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is a ChatGPT plugin?
&lt;/h2&gt;

&lt;p&gt;For OpenAI's setup now, each plugin sits in the directory &lt;a href="https://developers.openai.com/plugins/concepts/plugins" rel="noopener noreferrer"&gt;shared by ChatGPT and Codex&lt;/a&gt;. A plugin can have procedures, an MCP server, or both: procedures show the AI a repeatable workflow, while the MCP server provides current information, authentication, and limited tasks. A MCP-backed plugin may also include UI assets for places that run MCP Apps. The manifest lives at &lt;code&gt;.codex-plugin/plugin.json&lt;/code&gt;, with &lt;code&gt;skills/&lt;/code&gt;, &lt;code&gt;hooks/&lt;/code&gt;, and MCP configuration at the package root.&lt;/p&gt;

&lt;p&gt;It carries baggage. The phrase "ChatGPT plugin" also means OpenAI's 2023 OpenAPI manifests toolset, discontinued in 2024. They have only the name in common: the new setup uses MCP to handle tool access, while Agent Skills gives guidance, with one plugin from the current directory as those parts, rather than any OpenAPI wrapper.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do plugins, skills, connectors, and hooks fit together?
&lt;/h2&gt;

&lt;p&gt;The parts stack cleanly once the labels are split. A skill is workflow steps kept in the folder &lt;code&gt;SKILL.md&lt;/code&gt;. The backend an assistant reaches out to is the connector or MCP. A lifecycle handler runs at set times during a session, as described by &lt;a href="https://www.speakeasy.com/resources/ai-agent-hooks" rel="noopener noreferrer"&gt;AI agent hooks&lt;/a&gt;. A plugin bundles any mix of them with one label and one release, so they distribute and update together.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Harness&lt;/th&gt;
&lt;th&gt;What "plugin" means&lt;/th&gt;
&lt;th&gt;MCP role&lt;/th&gt;
&lt;th&gt;Extra components&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Cursor&lt;/td&gt;
&lt;td&gt;Agent Plugins (portable) and Cursor Plugins (native)&lt;/td&gt;
&lt;td&gt;Bundled &lt;code&gt;mcp.json&lt;/code&gt; server definitions&lt;/td&gt;
&lt;td&gt;Skills in both; Cursor Plugins add rules, agents, commands, hooks, variables&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude&lt;/td&gt;
&lt;td&gt;Installable package for Code, Cowork, and chat&lt;/td&gt;
&lt;td&gt;Connectors are MCP servers&lt;/td&gt;
&lt;td&gt;Skills, subagents, hooks; Code adds LSP servers and monitors&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;ChatGPT&lt;/td&gt;
&lt;td&gt;Package in the shared ChatGPT and Codex directory&lt;/td&gt;
&lt;td&gt;MCP server is an optional component&lt;/td&gt;
&lt;td&gt;Skills, optional UI resources, lifecycle hooks&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;In real terms, claiming "we support plugins" means very little unless you find out what a harness really runs. An add-on that runs end to end with Claude Code may load in a conversation with only its abilities on; for any Cursor Plugin's guidelines or hooks, nothing moves to another harness.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why do security teams care about agent plugins?
&lt;/h2&gt;

&lt;p&gt;One marketplace setup dropping a config onto each machine can also drop an unreviewed MCP service or plugin onto each machine.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;They are the fleet-distribution mechanism.&lt;/strong&gt; Hooks are the interface for governing coding agents from inside the agent, and a hook config on one laptop is worthless if the other 999 laptops lack it. Every major harness now has a required-install mode: Cursor's Required plugins, Claude's organization-required plugins, and managed settings that force-enable a marketplace. A required plugin is how &lt;a href="https://www.speakeasy.com/resources/ai-agent-hooks" rel="noopener noreferrer"&gt;agent hooks&lt;/a&gt; and observability reach every machine without a per-laptop setup step.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;They install local MCP servers.&lt;/strong&gt; A plugin's &lt;code&gt;mcp.json&lt;/code&gt; can define a STDIO server that runs on the laptop and never crosses the &lt;a href="https://www.speakeasy.com/resources/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt;, so a marketplace install can quietly widen the tool surface outside the governed path. That is the &lt;a href="https://www.speakeasy.com/resources/shadow-mcp" rel="noopener noreferrer"&gt;shadow MCP&lt;/a&gt; problem arriving through a sanctioned channel.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;They carry ungoverned instructions.&lt;/strong&gt; A plugin's skills arrive as markdown, fire no procurement event, and steer agent behavior on every matching task. &lt;a href="https://www.speakeasy.com/resources/skills-governance" rel="noopener noreferrer"&gt;Skills governance&lt;/a&gt; covers why that surface needs an owner, a review, and a revocation path of its own.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Marketplace checks can help, but coverage varies with each harness, while in-house marketplaces leave the work with the company. When checking MCP hosts, defenders must know the same thing about plugins: what's loaded, where, and holding which contents.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does the Speakeasy AI Control Plane work with agent plugins?
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://www.speakeasy.com/resources/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt; uses plugins as its distribution layer rather than shipping a custom installer. From &lt;a href="https://www.speakeasy.com/docs/ai-control-plane/mcp-gateway/plugins" rel="noopener noreferrer"&gt;Distribute &amp;gt; Plugins&lt;/a&gt;, an admin bundles MCP servers and skills into managed plugins, assigns them to roles, and publishes them to Claude Code, Cursor, and Codex marketplaces through a GitHub-backed marketplace the platform hosts on the org's behalf.&lt;/p&gt;

&lt;p&gt;Compatible plugins &lt;a href="https://www.speakeasy.com/blog/release-agent-plugins-support" rel="noopener noreferrer"&gt;publish in the Agent Plugins 1.0 format&lt;/a&gt;, so one portable package installs across agents that support the standard, with legacy per-harness manifests bundled for older releases. Credentials aren’t stored in bundles: portable &lt;code&gt;mcp.json&lt;/code&gt; lists only service URLs and connection kinds, while authentication runs inside the app via OAuth. Those hook configurations powering policy and tool logging are what the platform's observability plugin carries; since it ships through every harness's plugin setup, an admin can make it mandatory, and all machines get controls at the following run, using &lt;a href="https://www.speakeasy.com/resources/ai-agent-hooks" rel="noopener noreferrer"&gt;MDM&lt;/a&gt; instead.&lt;/p&gt;

&lt;p&gt;Use a single plugin release to distribute MCP know-how and rights across varied tools, then &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk with us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What is an agent plugin?
&lt;/h3&gt;

&lt;p&gt;An agent plugin is an installable package that extends an AI agent by bundling reusable capabilities: typically skills (workflow instructions), one or more Model Context Protocol (MCP) servers or connectors, and harness-specific extras such as hooks, rules, commands, or subagents. The point of the package is distribution. Instead of wiring a server config, a set of skills, and a hook file onto each machine by hand, a team publishes one plugin and people or admins install it from a marketplace or a repo.&lt;/p&gt;

&lt;h3&gt;
  
  
  How is an agent plugin different from an MCP server?
&lt;/h3&gt;

&lt;p&gt;An MCP server is a running program that exposes tools, resources, and prompts to an agent over the Model Context Protocol. An agent plugin is the package that delivers capabilities to the agent, and an MCP server is one thing it can deliver: the plugin's mcp.json (or a connector entry) tells the agent where the server is and how to reach it. The server is the tool backend, and the plugin is the distribution unit. A plugin can also ship no server at all and contain only skills.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is the Agent Plugins standard?
&lt;/h3&gt;

&lt;p&gt;Agent Plugins is an open, vendor-neutral specification for the plugin package, released as version 1.0 in August 2026 with core maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel. It fixes the layout: a plugin.json manifest at the root, an mcp.json file with MCP server definitions, and a skills/ directory in the Agent Skills format, with harness-specific extras namespaced on top. The spec leaves installation, distribution, marketplaces, permissions, and credential handling to each harness, which is why one conformant package installs across Cursor, Codex, ChatGPT, GitHub Copilot, Kiro, and VS Code while the install and governance experience still differs by harness.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is a Cursor plugin?
&lt;/h3&gt;

&lt;p&gt;Cursor supports two plugin formats. Plugins that follow the Agent Plugins open standard have a plugin.json manifest at the root and package skills and MCP servers, and they load in Cursor without changes. Cursor Plugins use a .cursor-plugin/plugin.json manifest and add Cursor-specific components: rules, agents, commands, hooks, and variables. Both install from the Cursor Marketplace or from team marketplaces, where admins set each plugin to Default Off, Default On, or Required.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is a Claude plugin?
&lt;/h3&gt;

&lt;p&gt;A Claude plugin is a directory that bundles skills, connectors (MCP servers), subagents, and hooks, identified by a .claude-plugin/plugin.json manifest. In Claude Code, plugins can also carry LSP servers, background monitors, and default settings, and they install from GitHub-backed marketplaces. In the Claude apps, plugins install from the Customize menu into chat and Claude Cowork; the bundled skills work everywhere, while hooks and subagents run only in Cowork and Claude Code. Team and Enterprise admins can mark a plugin as required for the whole organization.&lt;/p&gt;

&lt;h3&gt;
  
  
  What is a ChatGPT plugin?
&lt;/h3&gt;

&lt;p&gt;In the current OpenAI system, a plugin is the package people discover and install from the plugin directory that ChatGPT and Codex share. It can contain skills, an MCP server, or both, plus optional UI resources, under a .codex-plugin/plugin.json manifest. This is unrelated to the 2023 ChatGPT plugins built on OpenAPI manifests, which OpenAI retired in 2024. The modern system is built on MCP and the Agent Skills format.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do agent plugins relate to skills and hooks?
&lt;/h3&gt;

&lt;p&gt;Skills and hooks are components inside a plugin. A skill is a folder of workflow instructions in the Agent Skills format that teaches the agent how to do a repeatable job. A hook is a lifecycle handler that runs at defined points in a session, such as before a tool call. A plugin is the wrapper that carries them: it gives skills, hooks, MCP server configs, and other components one versioned identity, so installing the plugin installs all of them together and updating it updates them together.&lt;/p&gt;

&lt;h3&gt;
  
  
  How does Speakeasy use agent plugins?
&lt;/h3&gt;

&lt;p&gt;The Speakeasy AI Control Plane packages MCP servers and skills into managed plugins and publishes them to Claude Code, Cursor, and Codex marketplaces through a GitHub-backed marketplace. Compatible plugins publish in the Agent Plugins 1.0 open standard, so one portable package installs across agents that support it, and packages never contain credentials, since authentication happens in the agent through OAuth. Speakeasy also ships its hook configurations as native Claude Code and Cursor plugins, so an admin can mark the plugin as required and the whole fleet picks up observability and policy on the next session.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What are background agents?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Tue, 10 Nov 2026 11:37:01 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-are-background-agents-2g7h</link>
      <guid>https://dev.to/speakeasy/what-are-background-agents-2g7h</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/background-agents" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Background agent
&lt;/h2&gt;

&lt;p&gt;A background agent executes asynchronously or outside the foreground&lt;br&gt;
interaction. Background describes how work runs, not who the agent works for,&lt;br&gt;
whether a person remains involved, which authority it uses, or how&lt;br&gt;
independently it decides.&lt;/p&gt;

&lt;p&gt;Background agents are frequently called autonomous agents, but the terms aren't synonyms. Autonomous describes how independently an agent decides and acts. Background describes where or how execution happens. A nightly triage worker can be both. Background work can also stream progress while a person actively participates, and an unattended agent can pause at an approval gate.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://www.speakeasy.com/resources/agent-types" rel="noopener noreferrer"&gt;four types of AI agents&lt;/a&gt; classify accountability and engagement. When background work proceeds without active human participation, it is an &lt;strong&gt;unattended individual agent&lt;/strong&gt; if it completes work for one person, or an &lt;strong&gt;unattended organization agent&lt;/strong&gt; if it performs a standing job for the organization. The &lt;a href="https://www.speakeasy.com/resources/user-steered-agents" rel="noopener noreferrer"&gt;user-steered agent&lt;/a&gt; is typically an attended individual agent, even when it starts a background operation and streams the result.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/background-agents#:~:text=When%20background%20work%20proceeds%20without%20active%20human%20participation" rel="noopener noreferrer"&gt;View the background agent identity and access diagram on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do common background labels map to the four types?
&lt;/h2&gt;

&lt;p&gt;Workflow, product, and assistant-style labels describe triggers, architecture, or product experience. They do not define a working arrangement by themselves.&lt;/p&gt;

&lt;h3&gt;
  
  
  Workflow agents
&lt;/h3&gt;

&lt;p&gt;A workflow agent is deployed to run a specific workflow. A schedule, queue, or event starts each run. A nightly support-triage worker is usually an unattended organization agent. A scheduled research task for one person can instead be an unattended individual agent. The trigger alone does not decide accountability or authority.&lt;/p&gt;

&lt;h3&gt;
  
  
  Product agents
&lt;/h3&gt;

&lt;p&gt;A product agent is a product that many people use. That architecture can support different arrangements. A customer-support agent working with a customer in live chat is an attended organization agent. A vendor-hosted coding agent that prepares a pull request after its requester leaves is an unattended individual agent. Product identity does not determine the cell.&lt;/p&gt;

&lt;h3&gt;
  
  
  Assistant-style agents
&lt;/h3&gt;

&lt;p&gt;An organization can set up an assistant-style agent for many people to use. An internal Slack assistant is an attended organization agent when employees actively participate in each exchange. If a person delegates a task and returns later for the result, that run is unattended individual. Assistant describes the experience, not accountability, engagement, or authority. Each request must carry the requester's authority without giving the agent a standing copy of their credentials.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why can the same software use different working arrangements?
&lt;/h2&gt;

&lt;p&gt;Classify each run rather than assigning the software a permanent type. Cursor used to call its Cloud Agents &lt;a href="https://cursor.com/help/ai-features/background-agents" rel="noopener noreferrer"&gt;Background Agents&lt;/a&gt;, and GitHub calls its Copilot coding agent an &lt;a href="https://github.blog/changelog/2026-02-17-delegate-tasks-to-copilot-coding-agent-from-visual-studio/" rel="noopener noreferrer"&gt;asynchronous, autonomous background agent&lt;/a&gt;. In both products, a person assigns a task, leaves, and later reviews a pull request. The run is unattended individual because it proceeds without active participation but remains accountable to that person.&lt;/p&gt;

&lt;p&gt;Grok on X illustrates another combination. It has a product identity and persists between requests, but a run can be attended organization when a person actively exchanges messages with it. A coding agent launched from a terminal can be attended individual. A scheduled run can be unattended individual or unattended organization, depending on who the work is for.&lt;/p&gt;

&lt;p&gt;The vendor's label does not decide the working arrangement or identity strategy. Accountability asks who the work is for. Engagement asks whether a person actively participates. Authority separately asks whether each action uses delegated-user, service/workload, or hybrid authority.&lt;/p&gt;

&lt;h2&gt;
  
  
  What identity do background agents need?
&lt;/h2&gt;

&lt;p&gt;Identity follows the authority for the action, not whether execution happens in the background. An unattended individual agent can retain one person's delegated authority through a task-scoped credential. An unattended organization agent usually acts through a dedicated service or workload identity. Running either on the deployer's standing credentials or an undifferentiated shared bot credential creates the failures catalogued in &lt;a href="https://www.speakeasy.com/resources/non-human-identity" rel="noopener noreferrer"&gt;non-human identity&lt;/a&gt;. &lt;a href="https://www.speakeasy.com/resources/agent-identity" rel="noopener noreferrer"&gt;Agent identity&lt;/a&gt; explains what is bound, and &lt;a href="https://www.speakeasy.com/resources/agent-auth" rel="noopener noreferrer"&gt;agent auth&lt;/a&gt; explains how the agent proves that binding on each request.&lt;/p&gt;

&lt;h3&gt;
  
  
  A dedicated identity with a fixed grant
&lt;/h3&gt;

&lt;p&gt;Organization agents using service or workload authority require a dedicated account in the directory, assigned to a named human owner and granted credentials tailored to the agent's job the same way any other role would be approved. This applies to attended and unattended organization agents. The grant stays fixed across runs. A user session gets attribution, least privilege, and lifecycle handling automatically by being tied to a person in the directory. An organization agent has no person to inherit those controls from, so the system must provide them directly.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Attribution.&lt;/strong&gt; Every action resolves to the agent's own name, and through the owner of record to an accountable person, without being mistaken for that person's own activity.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Least privilege.&lt;/strong&gt; The grant covers the workflow the agent was deployed for, not the union of everything its deployer can reach.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lifecycle.&lt;/strong&gt; The agent does not stop when anyone leaves, so offboarding is an explicit event where the identity is reassigned to a new owner or revoked.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Just-in-time credentials per task
&lt;/h3&gt;

&lt;p&gt;An unattended individual agent has no person actively steering it, but each task still comes from a specific person and should carry only that person's authority. Giving the agent one standing grant broad enough to cover everyone who might ask it to act turns it into a confused deputy. Any requester could reach anything the agent can reach, regardless of what that requester is allowed to see.&lt;/p&gt;

&lt;p&gt;A &lt;a href="https://www.speakeasy.com/resources/task-scoped-credentials" rel="noopener noreferrer"&gt;task-scoped credential&lt;/a&gt; avoids that. The agent keeps a standing identity, including a name, an owner, and an inventory record, but no standing grant to anyone's data. Each request mints a short-lived credential that names the requesting person as the subject, the agent as the actor, and only the scope that task needs. The credential expires when the task finishes, and if the requesting person's access is later revoked, no new credential can be issued in their name. &lt;a href="https://www.speakeasy.com/resources/extend-iam-to-agents" rel="noopener noreferrer"&gt;Extending IAM to agents&lt;/a&gt; covers the same authority pattern for a custom workflow one person builds and another person triggers.&lt;/p&gt;

&lt;p&gt;When a vendor hosts product agents, that adds another boundary. The organization can't control Cursor's virtual machines or GitHub's runners. Therefore, the only identity it can manage is the one the agent shows when it accesses the organization's tools and data. That identity is the one to bind, scope, and revoke.&lt;/p&gt;

&lt;h2&gt;
  
  
  How should CISOs inventory background agents?
&lt;/h2&gt;

&lt;p&gt;Background agents belong in the &lt;a href="https://www.speakeasy.com/resources/agent-inventory" rel="noopener noreferrer"&gt;agent inventory&lt;/a&gt;, with fields matched to their type.&lt;/p&gt;

&lt;h3&gt;
  
  
  What should a background agent inventory record include?
&lt;/h3&gt;

&lt;p&gt;An unattended organization agent's record should include:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The identity it uses.&lt;/li&gt;
&lt;li&gt;The person or team responsible for it.&lt;/li&gt;
&lt;li&gt;The workflow it was set up for.&lt;/li&gt;
&lt;li&gt;The tools, data, and MCP servers it can access.&lt;/li&gt;
&lt;li&gt;Its trigger, such as a schedule, queue, or event.&lt;/li&gt;
&lt;li&gt;The last time it was active.&lt;/li&gt;
&lt;li&gt;The steps required to remove it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;An unattended individual agent's record also retains the requester, task boundary, delegated scope, and expiry. An assistant-style agent also tracks who can interact with it and the limits on what those requests can do.&lt;/p&gt;

&lt;p&gt;Discovery works differently than it does for the user-steered population. Endpoint and MDM tools can see Claude Desktop on a managed laptop, but they miss the worker that runs overnight in a virtual machine. Instead, background agents show up when you go to find tools.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;MCP gateway traffic&lt;/strong&gt; from clients that are not desktops or IDEs&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Service identities&lt;/strong&gt; in the identity provider with no matching human sign-in pattern&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Standing automations&lt;/strong&gt; and scheduled jobs in orchestration platforms&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bots that act on a schedule&lt;/strong&gt; with no corresponding human session&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each finding lands in the register as a record with an owner and a disposition, the same place that holds findings from &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;shadow AI&lt;/a&gt; discovery. An unauthorized background agent is still a principal that can access things, so it must be managed or removed from the record.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does the Speakeasy AI Control Plane govern background agents?
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://www.speakeasy.com/product/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt; supports both authority strategies on the path background agents take to tools. Organization agents using service or workload authority get a dedicated identity with a named human owner. Individual agents using delegated-user authority receive a temporary credential for each request, tied to the person and agent and deleted when the task finishes. Hybrid runs record the principal used for each action. The &lt;a href="https://www.speakeasy.com/product/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt; is the chokepoint between agents and the tools an organization exposes.&lt;/p&gt;

&lt;p&gt;This traffic keeps the inventory up to date. Every worker that accesses a tool through the gateway is logged as being used, showing a live "last-seen" time and &lt;a href="https://www.speakeasy.com/product/agent-observability" rel="noopener noreferrer"&gt;audit logs&lt;/a&gt; that the security team can download. When you remove someone's access, the process depends on what is being taken away. Switching ownership or canceling a unique identity is handled through the directory, while cutting off a person's account altogether stops their temporary credentials from being issued for any assistant they have used. To put an identity, an owner, and a revocation path behind the background agents already running in your organization, &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What are background agents?
&lt;/h3&gt;

&lt;p&gt;A background agent executes asynchronously or outside the foreground interaction. A request, schedule, queue, event, or system signal can start it. Background work is often unattended, but it can also stream progress while a person remains active in the session. Classify its working arrangement separately.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is a background agent the same as an autonomous agent?
&lt;/h3&gt;

&lt;p&gt;No. Autonomous describes how independently an agent decides and acts. Background describes where or how execution happens. An attended agent can run work in the background, and an unattended agent can pause for approval. Neither term identifies whether the run is accountable to an individual or the organization.&lt;/p&gt;

&lt;h3&gt;
  
  
  What are the types of background agents?
&lt;/h3&gt;

&lt;p&gt;Background is not a primary agent type. When the work proceeds without active human participation, it is either an unattended individual agent completing work for one person or an unattended organization agent performing a standing job. Workflow, product, and assistant-style describe triggers, architecture, or product experience and do not determine the working arrangement by themselves.&lt;/p&gt;

&lt;h3&gt;
  
  
  Are Cursor Cloud Agents and Copilot coding agent background agents?
&lt;/h3&gt;

&lt;p&gt;Yes, when a person assigns a task and leaves while the agent prepares the result. That run is an unattended individual agent because it remains accountable to the requester. Cursor once called its product Background Agents, and GitHub describes Copilot coding agent as an asynchronous, autonomous background agent. Those vendor labels describe execution and discretion, not the run's working arrangement or authority.&lt;/p&gt;

&lt;h3&gt;
  
  
  What identity do background agents need?
&lt;/h3&gt;

&lt;p&gt;Identity follows authority, not background execution. An unattended individual agent usually uses a task-scoped credential that preserves the requester's delegated authority. An unattended organization agent usually uses a dedicated service or workload identity with a named owner and a narrowly scoped grant. Hybrid runs must record which principal authorizes each consequential action.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do you find background agents you did not deploy?
&lt;/h3&gt;

&lt;p&gt;On the path to tools rather than on endpoints. Endpoint and MDM tooling sees Claude Desktop on a managed laptop and misses the worker that runs overnight in a VM. Background agents surface as MCP gateway traffic from clients that are not desktops or IDEs, service identities in the identity provider with no human sign-in pattern, standing automations in orchestration platforms, and bots that act on a schedule with no corresponding human session. Each finding lands in the agent inventory as a record with an owner and a disposition.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>infrastructure</category>
      <category>speakeasy</category>
    </item>
    <item>
      <title>What are user-steered agents?</title>
      <dc:creator>RitzaCo</dc:creator>
      <pubDate>Tue, 10 Nov 2026 08:10:00 +0000</pubDate>
      <link>https://dev.to/speakeasy/what-are-user-steered-agents-2pf1</link>
      <guid>https://dev.to/speakeasy/what-are-user-steered-agents-2pf1</guid>
      <description>&lt;p&gt;By &lt;a href="https://www.speakeasy.com/author/nolan-sullivan" rel="noopener noreferrer"&gt;Nolan Sullivan&lt;/a&gt;, Founding Growth Engineer at Speakeasy. &lt;a href="https://www.speakeasy.com/resources/user-steered-agents" rel="noopener noreferrer"&gt;Originally published on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  User-steered agent
&lt;/h2&gt;

&lt;p&gt;A user-steered agent is an attended individual agent that runs as a session&lt;br&gt;
attached to a person. The person opens it, gives it a task, watches it call&lt;br&gt;
tools, and decides each next step, as with Claude Desktop, Claude Code in a&lt;br&gt;
terminal, Cursor in the editor, and ChatGPT.&lt;/p&gt;

&lt;p&gt;User-steered agents run in a chat window, terminal, or editor while a person guides the session. The agent can make plans, use tools, and carry out multiple steps, while the user watches the process and chooses what to do afterward. A session can even &lt;a href="https://www.speakeasy.com/blog/release-agent-session-portability" rel="noopener noreferrer"&gt;move between Claude Code, Cursor, and Codex&lt;/a&gt; without changing type, because a person still steers it.&lt;/p&gt;

&lt;p&gt;In the four &lt;a href="https://www.speakeasy.com/resources/agent-types" rel="noopener noreferrer"&gt;types of AI agents&lt;/a&gt; defined by accountability and engagement, this is an &lt;strong&gt;attended individual agent&lt;/strong&gt;. User-steered describes the interaction pattern, while attended individual names the working arrangement.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do user-steered and background agents relate?
&lt;/h2&gt;

&lt;p&gt;A user-steered run is tied to one person's active session and is therefore attended and individual. Background is not its strict opposite. It describes where or how work executes, so a user-steered agent can start a background operation and remain attended while progress streams back. The run becomes unattended only when active participation is no longer required.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;User-steered run&lt;/th&gt;
&lt;th&gt;Background run&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Engagement&lt;/td&gt;
&lt;td&gt;Attended&lt;/td&gt;
&lt;td&gt;Often unattended, but can remain attended&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Accountability&lt;/td&gt;
&lt;td&gt;Individual&lt;/td&gt;
&lt;td&gt;Individual or organization&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Starts from&lt;/td&gt;
&lt;td&gt;A person's active session&lt;/td&gt;
&lt;td&gt;A request, schedule, queue, event, or system signal&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Authority&lt;/td&gt;
&lt;td&gt;Usually delegated-user authority&lt;/td&gt;
&lt;td&gt;Delegated-user, service/workload, or hybrid authority&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Example&lt;/td&gt;
&lt;td&gt;Claude Desktop, Claude Code in a terminal, Cursor in the IDE&lt;/td&gt;
&lt;td&gt;Delegated cloud coding or a nightly organizational workflow&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;An internal Slack agent can be an attended organization agent when employees actively exchange messages with it, even if the service persists between conversations. Product architecture and deployment lifetime do not determine attendance.&lt;/p&gt;

&lt;h2&gt;
  
  
  When does a user-steered agent stop being user-steered?
&lt;/h2&gt;

&lt;p&gt;Stopping steering is often simpler in theory than in practice. The same software acts differently depending on how it's started.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Claude Code in a terminal&lt;/strong&gt; is user-steered and attended individual. The same Claude Code launched by a scheduled trigger is unattended. It is individual if the schedule retains one person's delegated authority, or organization if it performs a standing organizational job.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cursor in the editor&lt;/strong&gt; is user-steered. Hand the task to a Cursor Cloud Agent and the person assigns it, walks away, and reviews a pull request.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Copilot in the IDE&lt;/strong&gt; is user-steered. Assign an issue to Copilot coding agent and the agent decides every step until the pull request is ready.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The delegated runs in the second and third cases are unattended individual agents. They still act on the assigning person's behalf, so the person's authority has to travel with the run without the live session. That is the case for a &lt;a href="https://www.speakeasy.com/resources/task-scoped-credentials" rel="noopener noreferrer"&gt;task-scoped credential&lt;/a&gt;, minted for the run, naming the person as subject and the agent as actor, and expiring when the pull request lands. A standing organizational schedule instead needs service or workload authority with a dedicated identity and owner.&lt;/p&gt;

&lt;h2&gt;
  
  
  What agent identity does a user-steered agent use?
&lt;/h2&gt;

&lt;p&gt;Usually the person's delegated identity. The session logs in using the person's single sign-on identity, each tool use is approved based on the person's permissions, and the audit log shows who was in control. This authority model is separate from attendance: an unattended individual agent can also retain delegated-user authority for a bounded task. &lt;a href="https://www.speakeasy.com/resources/agent-auth" rel="noopener noreferrer"&gt;Agent auth&lt;/a&gt; explains the OAuth mechanism that proves the binding to a resource, &lt;a href="https://www.speakeasy.com/resources/agent-identity" rel="noopener noreferrer"&gt;agent identity&lt;/a&gt; explains what is bound, and &lt;a href="https://www.speakeasy.com/resources/extend-iam-to-agents" rel="noopener noreferrer"&gt;extending IAM to agents&lt;/a&gt; shows how the existing user directory handles the setup.&lt;/p&gt;

&lt;p&gt;Three details keep inheritance precise:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://www.speakeasy.com/resources/user-steered-agents#what-agent-identity-does-a-user-steered-agent-use" rel="noopener noreferrer"&gt;View the user-steered agent identity diagram on Speakeasy&lt;/a&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Inheritance sets a ceiling.&lt;/strong&gt; The agent can never exceed the person's grants, and most organizations narrow it further, denying the agent destructive calls the person's own role allows.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The session records more than the person.&lt;/strong&gt; It binds the agent workload and the client too, so the same person driving Claude Code and Cursor produces two distinguishable actors. &lt;a href="https://www.speakeasy.com/resources/agent-composite-identity" rel="noopener noreferrer"&gt;Agent composite identity&lt;/a&gt; describes the tuple.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Offboarding follows the user account.&lt;/strong&gt; When the person is deprovisioned, every session bound to their identity loses access on the same directory event that disables their login.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How does a user-steered agent lose its governance advantage?
&lt;/h2&gt;

&lt;p&gt;This happens when the agent runs on something other than the person, often through a shared team bot token in an MCP configuration. The agent's actions are then attributed to the service account. Its access permissions become a combination of teammates' requirements and are not revoked when a team member leaves. Those are the machine-credential failures listed in &lt;a href="https://www.speakeasy.com/resources/non-human-identity" rel="noopener noreferrer"&gt;non-human identity&lt;/a&gt;. A user-steered agent should inherit the person's identity instead.&lt;/p&gt;

&lt;h2&gt;
  
  
  How should a CISO inventory user-steered agents?
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Count user-steered agent clients
&lt;/h3&gt;

&lt;p&gt;The agents are visible in the &lt;a href="https://www.speakeasy.com/resources/agent-inventory" rel="noopener noreferrer"&gt;agent inventory&lt;/a&gt;. Endpoint and MDM tooling detect Claude Desktop and Cursor installed on managed laptops. The identity provider tracks seats and sign-ins, and procurement tracks subscriptions. Counting user-steered agents is often straightforward.&lt;/p&gt;

&lt;h3&gt;
  
  
  Inventory the tools connected to each agent
&lt;/h3&gt;

&lt;p&gt;Their reach requires separate discovery. An MCP server connected to an editor can appear as a few lines in a dotfile, such as &lt;code&gt;~/.cursor/mcp.json&lt;/code&gt;, and a skill can be a markdown file in a project directory. An endpoint scan can detect the editor while overlooking its connections. &lt;a href="https://www.speakeasy.com/resources/shadow-ai" rel="noopener noreferrer"&gt;Shadow AI&lt;/a&gt; addresses this gap. A complete record tracks the person, client software, and connected MCP servers, skills, and tools. &lt;a href="https://www.speakeasy.com/resources/ai-agent-hooks" rel="noopener noreferrer"&gt;Agent hooks&lt;/a&gt; within the agent and gateway traffic surface those connections.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does the Speakeasy AI Control Plane govern user-steered agents?
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://www.speakeasy.com/product/ai-control-plane" rel="noopener noreferrer"&gt;Speakeasy AI Control Plane&lt;/a&gt; makes inheritance the enforced default for this type. Sessions log in using the identity provider your company already uses, meaning a user-steered agent accesses tools as the person using it, not through a general, shared key. The &lt;a href="https://www.speakeasy.com/product/mcp-gateway" rel="noopener noreferrer"&gt;MCP gateway&lt;/a&gt; checks each tool request against that person's permissions and logs it in &lt;a href="https://www.speakeasy.com/product/agent-observability" rel="noopener noreferrer"&gt;audit logs&lt;/a&gt; using their name. If the person is removed from the identity provider, their agents lose access at the same moment their login is disabled. When the same person hands a task to an unattended run, the platform switches strategy and creates a task-scoped credential for it. To bind the sessions your organization already runs to the people behind them, &lt;a href="https://www.speakeasy.com/talk-to-us" rel="noopener noreferrer"&gt;talk to us&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  What are user-steered agents?
&lt;/h3&gt;

&lt;p&gt;A user-steered agent, also called a user-based or interactive agent, is an attended individual agent that runs as a session attached to a person. The person opens it, gives it a task, watches it call tools, and decides each next step. Claude Desktop, Claude Code in a terminal, Cursor in the editor, and ChatGPT commonly work this way.&lt;/p&gt;

&lt;h3&gt;
  
  
  Is Claude Desktop a user-steered agent?
&lt;/h3&gt;

&lt;p&gt;Yes. A person opens Claude Desktop, gives it a task, watches it use tools, and decides what happens next, so the agent exists as that person's session and acts within that person's reach. Claude Code in a terminal, Cursor in the editor, and ChatGPT work the same way. When the person closes the session or is deprovisioned, the agent's access ends with them.&lt;/p&gt;

&lt;h3&gt;
  
  
  What identity does a user-steered agent use?
&lt;/h3&gt;

&lt;p&gt;It inherits the user's access. The session authenticates as the person through the organization's identity provider, every tool call is authorized against the person's grants, and the audit trail resolves to the person. Inheritance sets a ceiling rather than copying the grant, so most organizations restrict the agent to a subset of the person's roles. The session also records which agent and which client were used, so two agents driven by the same person stay distinguishable.&lt;/p&gt;

&lt;h3&gt;
  
  
  When does a user-steered agent stop being user-steered?
&lt;/h3&gt;

&lt;p&gt;When the person stops participating in the run. Assign an issue to Copilot coding agent or hand a task to a Cursor Cloud Agent, and the working arrangement becomes unattended individual while the run still acts for the requester. A scheduled run can be unattended individual when it retains one person's delegated authority, or unattended organization when it performs a standing organizational job.&lt;/p&gt;

&lt;h3&gt;
  
  
  Can a user-steered agent run on a shared service account instead?
&lt;/h3&gt;

&lt;p&gt;It can, and it is the most common way this type loses its governance advantage. A shared bot token in an MCP config means every action reads as the service account, the grant is the union of every teammate's needs, and no directory event ends the access when a person leaves. The session had a person to inherit and inherited a key instead.&lt;/p&gt;

&lt;h3&gt;
  
  
  How do you inventory user-steered agents?
&lt;/h3&gt;

&lt;p&gt;The agents themselves are easy to count: endpoint and MDM tooling sees Claude Desktop and Cursor installed, and the identity provider sees the seats. What that view misses is reach, because an MCP server or a skill wired into an editor is a few lines in a dotfile, not installed software. An inventory record for a user-steered agent binds the person, the client, and the MCP servers and tools connected to it, with the connections surfaced by hooks inside the agent or by the gateway traffic the sessions generate.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>mcp</category>
      <category>speakeasy</category>
      <category>infrastructure</category>
    </item>
  </channel>
</rss>
