DEV Community

Cover image for Agent harness or MCP gateway? What a company needs beyond the model, with AnythingMCP and 7 alternatives compared
helpcode.ai
helpcode.ai

Posted on

Agent harness or MCP gateway? What a company needs beyond the model, with AnythingMCP and 7 alternatives compared

In short

  • The model is the replaceable part. What decides whether AI is useful in a company is the software around it: the agent harness and the MCP gateway.
  • This article defines both, lists eleven things to check, and compares eight options against them.
  • We build AnythingMCP, one of the eight. Every statement about another product comes from that vendor's own pages, linked below, as read on 17 November 2026.

The model is the part you can swap

About one company in five now uses AI. In the European Union, 20.0% of enterprises with 10 or more employees used AI technologies in 2025, up from 13.5% a year earlier (Eurostat, 11 December 2025). In the United States, 19.8% of businesses used AI in a business function as of 3 May 2026, and 37% of firms with at least 250 employees did (U.S. Census Bureau, Business Trends and Outlook Survey, 26 May 2026).

Two figures side by side: 20.0% of EU enterprises with 10 or more employees used AI in 2025, and 19.8% of US businesses used AI in May 2026, 37% among firms with 250 or more employees

Most of that use starts the same way. Someone opens Claude, ChatGPT or Copilot and asks a question. The answer is good until the question needs something the model cannot see: an order in the ERP, a price list on the file server, a rule the company agreed years ago.

A better model does not fix that. The model only reads what it is given. Everything else is software around the model, and that software has two names you will meet in every product page this year: agent harness and MCP gateway.

What an agent harness is

An agent harness is the runtime around a model. It decides when to call the model, what to send it, how to read the answer, and which tools to call next.

Microsoft's documentation for Copilot Studio puts it this way: "The harness is a runtime that exists between the two: it determines when to call the model, what components to send it, interprets what comes back, and calls the right tools" (Harnesses in Copilot Studio). The same page lists what depends on the harness: creating and editing files, using skills, remembering context and orchestrating tool calls.

So when you use Claude, ChatGPT, Copilot or Cursor, you are using a model inside a harness. You usually do not choose the harness separately. It comes with the AI client.

Harness engineering is the work of designing that runtime: which context the model gets, which tools, which checks before and after each step.

What an MCP gateway is

MCP, the Model Context Protocol, is the open standard that lets a harness call tools. An MCP server offers tools. An MCP client, the harness, calls them.

An MCP gateway is one MCP endpoint in front of many tools. It handles sign-in, access control and the record of calls for all of them. AI clients connect to the gateway and not to each system.

Four layers: the model, the agent harness that runs the loop and calls tools, the MCP gateway that checks identity and roles and records each call, and the company systems behind it

The two are not alternatives. The harness sits with the model and belongs to the AI client. The gateway sits with your systems and belongs to you.

That split matters once a company uses more than one AI client, and most do. Memory, skills and file access are usually counted as harness features. If they live only in one harness, they stay in one AI client. If the gateway serves them, every harness that connects gets the same ones.

Eleven things to check beyond the model

A checklist in five groups: reach (connectors, file access), control (identity, roles per tool, audit log), knowledge (memory, skills), execution (code in a sandbox) and terms (hosting, licence, price model)

Reach

  1. Connectors for the systems you run. Count the ones you have, not the size of the catalog. Check the old ones too: a SOAP service, an OData service on SAP, a SQL database that is the real source of truth.
  2. File access. Much of what a company knows is in PDFs and spreadsheets on a file server. Ask whether the AI can search them and say which file, page and date an answer comes from.

A Claude answer from demo documents that quotes a price row and names the file, the page and the validity date of two price lists

Control

  1. Identity. Single sign-on from your directory, and per-user sign-in to the connected systems, so that each person sees what they are allowed to see there.
  2. Roles per tool. Start read-only. The role should decide which tools a person's AI can call.
  3. Audit log. Who called which tool, with which input, and what came back.

Knowledge

  1. Memory. A fact said once in one AI client should be known in the others.
  2. Skills. The way your company does a piece of work, written down as a procedure the AI follows.

Execution

  1. Code in a sandbox. One short program that calls several tools is cheaper and more exact than twenty single calls. It must run isolated, with no network and no secrets.

Terms

  1. Hosting. Your own servers, or a region you choose.
  2. Licence. Whether you can read the code and keep running it.
  3. Price model. Flat, per seat, or per tool call. An agent conversation can make dozens of calls.

Eight options compared

We picked the options a buyer would put on a shortlist today: two gateways from the API and AI infrastructure side, three integration platforms that added MCP, one agent builder with its own harness, one tool platform for agents, and our own.

A matrix of eight products against seven criteria: self-hosted, your own APIs as tools, roles and audit log, skills, memory, file search and code sandbox, with marks for stated, partly, not available and not stated

How to read the marks: a dot means the vendor's page did not say. It does not mean the feature is missing. The three tables below name the plan behind each mark.

Hosting, licence and price

Option What it is Where it runs Licence Price model
AnythingMCP MCP gateway that also generates the tools from your APIs Your servers with Docker, or AnythingMCP Cloud in Frankfurt AGPL-3.0; some modules under a commercial licence Self-hosted Community free up to 3 active users; Cloud €19 or €49 a month, flat; Enterprise quoted
Composio Tool platform for agents, with an MCP gateway Composio Cloud; bring your own cloud on Enterprise SDK repository MIT; platform not stated Free with 10,000 tool calls a month; Standard $18 per seat a month; extra calls $0.005 each; Enterprise custom
Zapier MCP Hosted MCP server over Zapier's apps Zapier's cloud, data in AWS US-East 1; no VPC or on-premises Not stated Runs on your existing Zapier plan with no separate MCP billing; each successful tool call uses two tasks
n8n Workflow automation; workflows you enable become MCP tools Self-hosted, or n8n Cloud Sustainable Use License; Enterprise files under a separate licence Self-hosted Community free; Cloud from €20 a month for 2,500 executions; Business €667 a month; Enterprise quoted
Workato Integration platform with Enterprise MCP Workato's cloud; a dedicated AWS VPC as Virtual Private Workato Not stated Quote only; usage tracked by tasks, on-premises agents and concurrency
Microsoft Copilot Studio Agent builder with its own harnesses and an MCP client Microsoft's service; self-hosting not stated Not stated Copilot Credits: $200 a month per 25,000 credits, or pay as you go
Kong AI Gateway API gateway with MCP plugins On-premises, or Kong Konnect MCP plugin only in the AI Gateway Enterprise offering Konnect Plus from $25 a month plus usage; Enterprise and self-hosted quoted
TrueFoundry AI gateway with an MCP gateway and registry SaaS; VPC, on-premises and air-gapped on Enterprise Not stated Free for up to 3 users; Pro $25 per user a month; Enterprise custom

Reach and control

Option Your own APIs as tools Roles and audit log Identity
AnythingMCP 800+ ready connectors; REST/OpenAPI, SOAP/WSDL, GraphQL, OData and SQL sources become tools without code Roles, per-tool permissions and an audit log in every edition SSO and SCIM on self-hosted Enterprise; per-user sign-in on Enterprise
Composio 1,700+ apps; custom connectors for internal apps on Enterprise Roles and tool scoping; audit logs on Enterprise SSO over SAML and OIDC on Standard and Enterprise; SCIM on Enterprise
Zapier MCP 9,000+ apps; your own APIs not stated App and action restrictions set for the account; a history of tool calls and an audit log SAML SSO on Enterprise
n8n Each tool is a workflow you build and enable for MCP Custom roles and audit logging on Enterprise SSO and LDAP on Business and Enterprise
Workato 12,000+ apps; recipe functions and skills become MCP servers Role-based access control and audit trails Agent actions inherit the signed-in user's identity
Microsoft Copilot Studio Power Platform connectors, custom connectors and MCP servers Power Platform data policies also govern MCP servers; audit not stated OAuth 2.0 per user for MCP servers
Kong AI Gateway Any Kong-managed API becomes MCP tools from its schema Access lists per tool, audit logs and metrics OAuth 2.0 plugin for MCP
TrueFoundry Registered MCP servers; internal APIs through custom servers Role-based access per MCP server; audit logging OAuth 2.0; SSO with Okta, Azure AD and others

Knowledge and execution

Option Skills Memory File search Code sandbox
AnythingMCP AI skills in every edition; learning projects on Team and Enterprise Company memory shared by every AI client, on Enterprise SFTP, FTP, WebDAV, S3 and network shares; on Cloud, or self-hosted Enterprise Code Mode in an isolated runner, switched on per MCP server
Composio "Skill sharing" listed in the plans; details not stated Not stated Not stated Sandboxes "to run code and process files"
Zapier MCP Not stated Not stated Not stated Not stated
n8n Not stated Not stated Not stated Not stated
Workato "Enterprise Skills": recipe functions served as MCP servers Not stated Not stated Not stated
Microsoft Copilot Studio Skills on two of its three harnesses Memory on the same two harnesses, for agents built in Copilot Studio Creates and edits Word, Excel, PowerPoint and PDF files; file server search not stated Each task runs in a sandbox on the Microsoft Copilot harness
Kong AI Gateway Not stated Not stated Not stated Not stated
TrueFoundry Agent Skills Registry, a separate product Not stated Not stated Agents can "generate and execute code"; sandboxing not described

Sources, all read on 17 November 2026:

Where each option fits best

No option is the best fit for every company. Read the tables with your own situation in mind.

  • AnythingMCP fits a company with more than one AI client and with systems that have no MCP server of their own: an ERP, SOAP or OData services, SQL databases, file servers. It generates the tools from those sources and serves skills, memory and file search from the gateway, so Claude, ChatGPT and Copilot all get the same ones. We did not find that combination stated by any other vendor above.
  • Zapier MCP has the largest catalog, 9,000+ apps, and nothing to run. For standard SaaS actions and light use it is the shortest path.
  • Composio covers 1,700+ apps and includes sandboxes that run code. Its paid plans add a firewall and zero data retention.
  • Workato suits a company that already runs it as its integration platform. MCP is then an extension of what is there.
  • Microsoft Copilot Studio brings skills, memory, file editing and a sandbox inside its own harness. For a company that works in Microsoft 365 and stays there, that is a strong reason.
  • Kong AI Gateway fits a company that already publishes its APIs through Kong. The same gateway then serves them as MCP tools.
  • n8n is the better engine when the AI is one step in a longer workflow with branches, schedules and error handling.
  • TrueFoundry counts model requests and MCP tool calls on one platform and offers air-gapped deployment on Enterprise.

A 10-minute self-hosted start

This is the quickstart from the AnythingMCP README. It needs Docker 24 or newer and openssl. On macOS, start Docker Desktop first.

mkdir anythingmcp && cd anythingmcp
curl -fsSLo docker-compose.yml \
  https://raw.githubusercontent.com/HelpCode-ai/anythingmcp/main/docker-compose.quickstart.yml
printf 'JWT_SECRET=%s\nENCRYPTION_KEY=%s\n' "$(openssl rand -hex 32)" "$(openssl rand -hex 32)" > .env
docker compose up -d
Enter fullscreen mode Exit fullscreen mode

Open http://localhost:3000 and register. The first account becomes admin. Your MCP endpoint is http://localhost:4000/mcp, and the API documentation is at http://localhost:4000/api/docs.

Keep the generated .env file. ENCRYPTION_KEY decrypts the credentials you store. If you lose it, every connector needs its credentials again.

Then three steps:

  1. Add a connector. Install one from the catalog, or point AnythingMCP at your own OpenAPI spec, WSDL, GraphQL endpoint, Postman collection or database. Each operation becomes an MCP tool.
  2. Decide what the model sees. Rename and describe the tools, drop the fields that must not leave your network, and choose which roles may call which tools.
  3. Hand one URL to your AI client. An MCP server in AnythingMCP is the endpoint you add to Claude, ChatGPT, Copilot, Gemini or Cursor. It exposes the connectors you assign to it and nothing else.

The tool list of a SAP S/4HANA connector in AnythingMCP: ten MCP tools, each with Test, Edit, Hints, Disable and Delete

The quickstart listens on 127.0.0.1 only, because nothing in front of it handles TLS. For an instance that colleagues or a cloud AI client can reach, clone the repository and run ./setup.sh. It asks for a domain, gets certificates and generates the secrets. The deployment guide has the details.

How to choose

Your situation Look first at
Several AI clients, and systems with no MCP server: ERP, SOAP, databases, file servers AnythingMCP
Everything must run on your own servers at no licence cost AnythingMCP Community or n8n Community
Standard SaaS apps, light use, nothing to operate Zapier MCP
A large SaaS catalog, priced per seat Composio
You already run Workato, Kong or n8n The MCP feature of the platform you have
Your company works in Microsoft 365 and will stay there Copilot Studio, with a gateway for the systems its connectors do not cover
You also need to govern model traffic and cost TrueFoundry or Kong AI Gateway

A gateway and a harness can be combined. Copilot Studio and the common AI clients can call an MCP gateway as one of their tool sources.

Questions people ask

What is an agent harness?
The runtime around a model. It decides when to call the model, what context to send, and which tools to call with the answer. Claude, ChatGPT, Copilot and Cursor each contain one.

What is an MCP gateway?
One MCP endpoint in front of many tools. It handles sign-in, access control and the record of calls, so AI clients connect to the gateway and not to each system.

Agent harness or MCP gateway: do I need both?
You already have a harness if you use an AI client. You need a gateway when more than one person or more than one AI client should reach your systems under rules you set.

What is harness engineering?
The work of designing a harness: which context the model gets, which tools it can call, and which checks run before and after each step.

My AI client already has connectors. Why add a gateway?
Built-in connectors cover well-known SaaS apps and live in one AI client. A gateway reaches the systems that have no ready connector, and gives every AI client the same tools, roles and log.

Which of these can I self-host for free?
AnythingMCP states a free Community edition for up to 3 active users, and n8n a free Community edition. Composio, Kong and TrueFoundry tie self-hosting to an Enterprise arrangement. Zapier says it is not available.

Is AnythingMCP open source?
Yes, under AGPL-3.0. Some modules, among them the File Search sync engine and learning projects, are under a commercial licence and can be read in the same repository.

Where to go next

If you compare options yourself, take the eleven points above to each vendor and ask for the page where each one is stated.

Top comments (0)