MCP server tools now in ChatGPT -- developer mode

Coding with ChatGPT
1

ChatGPT developer mode is a beta feature that provides full Model Context Protocol (MCP) client support for all tools, both read and write. How to get started: https://platform.openai.com/docs/guides/developer-mode

Quick video on how it works:

11 Likes
3

Tried using it. The tools are loading, but when the model tries to invoke tools I get Http 424 errors.

ToolError: UNKNOWN: Error code: UNKNOWNError: HTTPException: 424: unhandled errors in a TaskGroup (1 sub-exception)

Looking at the logs I see the client disconnects midway through a call.

info: ModelContextProtocol.Server.McpServer[2103773711]
Server (McpPlayground.Mcp.Server 1.0.0.0), Client (openai-mcp 1.0.0) message processing canceled.
info: ModelContextProtocol.Server.McpServer[2081236114]
Server (McpPlayground.Mcp.Server 1.0.0.0), Client (openai-mcp 1.0.0) shutting down.
info: ModelContextProtocol.Server.McpServer[613617602]
Server (McpPlayground.Mcp.Server 1.0.0.0), Client (openai-mcp 1.0.0) shut down.

Regardless, good move on adding support for the full toolset.

2 Likes
4

Update for anyone running to the same problem

Had to revert the return types of tools from complex structured objects to strings.

Quite weird, since Claude Web had no issues with it.

2 Likes
5

This is super exciting! :tada:

From what I can tell, ChatGPT in developer mode only supports MCP servers that use OAuth or no authentication.

But what if the MCP server I want to connect to only supports a bearer token? How should I configure that in ChatGPT?

Here’s the setup I have:

{
  "mcpServers": {
    "observe": {
      "command": "npx",
      "args": [
        "mcp-remote@latest",
        "https://132284333042.observe-eng.com/v1/ai/mcp",
        "--header",
        "Authorization:${AUTH_HEADER}"
      ],
      "env": {
        "AUTH_HEADER": "Bearer 132284333042 f86wKuxTvEOSZ4X1I1-redacted"
      }
    }
  }
}
3 Likes
6

Great! When do we expect this support to be rolled out to Team workspaces? As of now we’re limited to search/fetch tools for custom connectors.

1 Like
7

Does this not work for existing chats for anyone else? The extra tools seem to be greyed out unless you’re starting a fresh chat.

8

It works with oauth: allthough In my case, one of my oauth scopes didnt have the refresh token grant, matching the client grants. Which claude.ai weirdly did not complain about but chatgpt did. Allthough that probably explains why my connection got lost in claude.ai after a while :sweat_smile:

9

As @mucore mentioned earlier, it seems that the MCP bridge used in chatgpt requires the arguments to be sent as a string, not objects. :confused: Didnt have that issue anywhere else (claude, mcp inspector).

10

It doesn’t work in project chats. I wasted a few hours trying to figure this out.

11

Also, it seems chatgpt requires specific tools like fetch and search:
https://platform.openai.com/docs/mcp#create-an-mcp-server

1 Like
12

Great to see this support added!

It is failing 99% of the time though to call my MCP server. The list_resources call correctly finds the tools, but the response is almost always “The xyz tool isn’t available”, even though it absolutely does show in the list.

Very occasionally it works by being very explicit (call xyz tool on MyConnector) but mostly even that fails. It calls the tool but returns an error ResourceNotFound: not found: MyConnector/link_68c2f524b0088191beefa6b03ac93cf5/xyz

Is this just me or a common issue?

13

@askibinski That confused me as well, but I figured out that while Deep Research requires those search and fetch endpoints (https://platform.openai.com/docs/mcp), Developer Mode can use any tool the MCP provides. (https://platform.openai.com/docs/guides/developer-mode)

@edwinarbus However, the problem I’m seeing is that it doesn’t look like file handling works in Developer Mode. When a tool call to my MCP server returns a file, ChatGPT thinks that it’s stored the file at the local path and tries to access it there, but fails. I’m not sure if the issue is that it failed to store the file, if it never stored it at all and is just hallucinating a filesystem path, or something else. My MCP also returns the same file at a publicly accessible URL, but ChatGPT can’t seem to fetch the file from the URL either. I can provide more info if someone on your team wants to debug this with me.

14

(post deleted by author)

15

One observation is that things work better with fewer tools. Our server has quite a few (70+). Disabling about half of them made it work significantly better.

Will test to see how it deals with multiple servers – if the limit is 30-40 tools total that would limit usability quite a bit

16

I am glad to see these added, but this is one area where Claude has got ChatGPT beaten hands down. With Claude, I can easily connect both remote and local MCP servers and do things like access my file system if I want to and other things like my reminders and other apps on my computer.

I really would like to see the ability to add local MCP servers to ChatGPT so that I can do things like query my Obsidian vault directly in ChatGPT and so forth.

1 Like