<?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: Privometry Team</title>
    <description>The latest articles on DEV Community by Privometry Team (@privometry).</description>
    <link>https://dev.to/privometry</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%2F4240422%2F2f488b99-8d0c-441d-9363-e2bb53fef979.png</url>
      <title>DEV Community: Privometry Team</title>
      <link>https://dev.to/privometry</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/privometry"/>
    <language>en</language>
    <item>
      <title>We Built an MCP Gateway to Control What AI Agents Can Actually Do</title>
      <dc:creator>Privometry Team</dc:creator>
      <pubDate>Thu, 12 Nov 2026 13:01:23 +0000</pubDate>
      <link>https://dev.to/privometry-org/we-built-an-mcp-gateway-to-control-what-ai-agents-can-actually-do-36oe</link>
      <guid>https://dev.to/privometry-org/we-built-an-mcp-gateway-to-control-what-ai-agents-can-actually-do-36oe</guid>
      <description>&lt;p&gt;MCP is making it much easier to connect AI assistants and agents to real tools.&lt;/p&gt;

&lt;p&gt;A coding assistant can read GitHub issues, interact with Jira, query databases, or access internal services through an MCP server.&lt;/p&gt;

&lt;p&gt;That's useful. But once an AI agent can call those tools, a different set of questions becomes important:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Which agent is making the request, and which person is responsible for it?&lt;/li&gt;
&lt;li&gt;Can the agent read information without being allowed to modify or delete it?&lt;/li&gt;
&lt;li&gt;What happens if an MCP server changes a tool's definition after it was approved?&lt;/li&gt;
&lt;li&gt;Can we require human approval before a sensitive action?&lt;/li&gt;
&lt;li&gt;How do we investigate what an agent attempted and what was actually executed?&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are the problems we're building &lt;strong&gt;Privometry MCP Security&lt;/strong&gt; to address.&lt;/p&gt;

&lt;h2&gt;
  
  
  The idea: put an authorization boundary between agents and tools
&lt;/h2&gt;

&lt;p&gt;Privometry is a security and governance gateway for MCP traffic.&lt;/p&gt;

&lt;p&gt;Instead of connecting supported AI clients directly to upstream MCP servers, teams can route those connections through Privometry.&lt;/p&gt;

&lt;p&gt;The gateway evaluates the request before forwarding it.&lt;/p&gt;

&lt;p&gt;The basic flow looks like this:&lt;/p&gt;

&lt;p&gt;&lt;code&gt;AI Agent → Privometry MCP Gateway → MCP Server → Enterprise System&lt;/code&gt;&lt;/p&gt;

&lt;p&gt;Inside the gateway, we focus on five areas.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Agent identity and access&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Track registered agents, their owners, their identities, and the people they act on behalf of. This provides context for authorization rules instead of relying on one shared credential.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Trusted MCP tools&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Review and approve servers and tools. Record trusted tool definitions, detect changes, and require re-review when a previously approved definition changes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Policy-based authorization&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Use Cedar policies to decide which agents can invoke which tools. The gateway can allow an action, deny it, or require human approval through Privometry's policy workflow.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Request and response protection&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Inspect tool arguments and responses for sensitive credentials and suspicious content, including potential prompt-injection attempts.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Activity and audit visibility&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Record tool calls, identities, authorization decisions, and execution outcomes so teams can investigate agent activity.&lt;/p&gt;

&lt;h2&gt;
  
  
  An example: preventing repository deletion
&lt;/h2&gt;

&lt;p&gt;Suppose a team connects an AI coding assistant to a GitHub MCP server.&lt;/p&gt;

&lt;p&gt;The assistant might legitimately need to search issues, inspect repositories, or create development tickets. That doesn't mean it should be allowed to delete repositories.&lt;/p&gt;

&lt;p&gt;In Privometry, a Cedar rule can explicitly deny a sensitive tool:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;@id("no-repo-deletes")
@verdict("deny")
@message("Agents may not delete repositories.")
forbid(
  principal,
  action == Mcp::Action::"call_tool",
  resource == Mcp::Tool::"github/delete_repo"
);
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This uses Privometry's MCP Cedar schema and rule annotations.&lt;/p&gt;

&lt;p&gt;The gateway evaluates the rule before forwarding the invocation. If the rule denies the action, the upstream MCP tool is not called.&lt;/p&gt;

&lt;p&gt;We also support approval-oriented policies, where sensitive operations can be held for human review.&lt;/p&gt;

&lt;p&gt;The intention is to make authorization &lt;strong&gt;deterministic and independently enforced&lt;/strong&gt;, rather than asking an AI model to decide whether it should obey a security rule.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why not rely entirely on MCP server authentication?
&lt;/h2&gt;

&lt;p&gt;Authentication and authorization at the underlying service are essential. Privometry is not intended to replace them.&lt;/p&gt;

&lt;p&gt;But an upstream service may know that a credential is valid without knowing which agent initiated the operation, what governance policy applies to that agent, or whether an additional human approval is required.&lt;/p&gt;

&lt;p&gt;A gateway adds another enforcement boundary, with centralized policies and activity visibility.&lt;/p&gt;

&lt;p&gt;There is an important limitation: &lt;strong&gt;a gateway governs traffic that actually passes through it&lt;/strong&gt;. Direct connections that bypass the gateway must be identified and addressed separately.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we're building toward
&lt;/h2&gt;

&lt;p&gt;Our focus is deliberately narrow for the first product: MCP security, identity-aware tool access, policy enforcement, approvals, and visibility.&lt;/p&gt;

&lt;p&gt;We're also working toward a self-service experience with a usage-limited free plan, reusable security policy templates, and more accessible policy authoring.&lt;/p&gt;

&lt;p&gt;We're interested in feedback from developers, platform engineers, and security teams already deploying MCP connections.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How are you currently governing MCP tool access in your organization?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Are you relying on individual MCP servers, gateway policies, existing IAM systems, or a combination?&lt;/p&gt;

&lt;p&gt;We'd be interested in hearing what works and where the difficult parts remain.&lt;/p&gt;

&lt;p&gt;Learn more about the project: &lt;a href="https://www.privometry.com/" rel="noopener noreferrer"&gt;Privometry MCP Security&lt;/a&gt;&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Privometry is building security and governance capabilities for AI-to-tool connections. This article focuses only on MCP Security.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>ai</category>
      <category>mcp</category>
      <category>security</category>
    </item>
  </channel>
</rss>
