Skip to content

Does ChatGPT support MCP?

Yes, through custom connectors in Developer Mode on paid plans, for servers reachable at a URL. How it works, what it cannot do, and how it compares with Claude and Cursor.

By Catherine Williams-TreloarSep 2026·4 min read

Yes. ChatGPT supports MCP through custom connectors: on a paid plan, with Developer Mode enabled, you add an MCP server by its URL and its tools become available in the conversation. The server has to be reachable at a URL (a remote MCP server) because ChatGPT cannot launch a process on your machine the way Claude Desktop or Cursor can.

That one constraint decides most of what follows. If the server you want is remote, ChatGPT is a full MCP client for it. If it is local-only, it is not reachable from ChatGPT until it is hosted.

ChatGPTbrowser or appBY URLRemote MCP serverStreamable HTTP · OAuthChatGPTbrowser or appNO STDIOLocal process on your machineno URL, so not reachable
Fig. 01: ChatGPT reaches servers by URL, and nothing else

How to add an MCP server to ChatGPT

In ChatGPT: Settings → Connectors → Advanced → Developer Mode, then Add custom connector. Paste the server's URL and give it a name. The connector is then enabled per conversation from the composer's connector menu.

For Annsa the URL is:

https://app.annsa.ai/mcp

The first use opens a browser on Annsa's consent screen. Approve once and the 6 tools appear. There is no API key to paste and no token file; access is your Annsa login, scoped to your workspace. Step-by-step setup is in connect Annsa to ChatGPT.

What ChatGPT needs from the server

ChatGPT expects the standard remote shape, which is the same shape Claude Code and Cursor expect:

  • Streamable HTTP: a POST endpoint that takes a JSON-RPC message and returns one
  • OAuth discovery: a 401 on the first request that points at protected-resource metadata, which points at authorization-server metadata
  • Dynamic client registration: so ChatGPT can register itself without a pre-issued client id
  • Authorization code with PKCE: the user approves in a browser; ChatGPT holds the token

A server built for Claude Code or Cursor over HTTP already meets all four. If it was built as a local stdio server, it needs the migration described in what is a remote MCP server and the auth chain in how to add OAuth to an MCP server.

What ChatGPT cannot do with MCP

  • Run a local server. No stdio transport, so no server that lives as a command on your laptop.
  • Read your files or your repo through the connector. The server can only return what it holds. Annsa's specs name files from a connected GitHub repository, but that connection is Annsa's, not ChatGPT's.
  • Enable a connector for every chat by default. Connectors are switched on per conversation.

None of these are limits on MCP. They are limits on where ChatGPT runs, which is a browser tab rather than your machine.

How it compares with Claude and Cursor

ChatGPTClaude CodeCursor
Remote servers by URLYes, custom connectorYes, claude mcp add --transport httpYes, Settings → MCP
Local stdio serversNoYesYes
Where it runsBrowser or appYour terminalYour editor
Can act on your code with the resultNoYesYes

For product management (asking what customers are saying, pulling the ranked priorities, reading a spec) ChatGPT with a remote connector is as capable as the others. For building the thing the spec describes, the editor-based clients are where the work happens. Setup for each is in using Annsa with coding tools.

Does ChatGPT support Model Context Protocol on the free plan?

Custom connectors sit behind Developer Mode, which is available on paid plans. On a free plan you can use the connectors OpenAI ships, but you cannot add your own server by URL.

The short answer, again

ChatGPT supports MCP for remote servers, added as custom connectors in Developer Mode on a paid plan, authenticated with OAuth in the browser. If your server has a URL and the standard auth chain, it works from ChatGPT with no changes.

annsa

Feedback in. Specs out.

500 extra pieces of feedback in your first month.