OpenAI API Compatibility Issue: Missing `additionalProperties: false` in Tool Parameter Schemas

[p-breslin]

Describe the bug

The github-mcp-server is incompatible with OpenAI models (like gpt-4.1 and gpt-4o-mini) that use function/tool calling. The OpenAI API requires tool parameter schemas to explicitly include "additionalProperties": false at the top level. Most tools generated by github-mcp-server are missing this property in their schema, causing API requests to fail with a 400 Bad Request error. This prevents the server from being used effectively with OpenAI-based clients.

Affected version

GitHub MCP Server
Version: v0.2.1
Commit: 9fa582d8d63522d70ce8f3af58265effb9645323
Build Date: 2025-04-21T23:03:01Z

Steps to reproduce the behavior

1. Run the ghcr.io/github/github-mcp-server:v0.2.1 Docker image (or build from source commit 9fa582d) with a valid GITHUB_PERSONAL_ACCESS_TOKEN.
2. Configure an MCP client to use this server with an OpenAI model that supports function/tool calling (e.g., gpt-4.1-nano).
3. Attempt to invoke a tool like add_issue_comment or get_pull_request through the client.
4. Observe the error returned by the OpenAI API (either in server logs or client response).

Expected vs actual behavior

Expected: The tool call should be successfully sent to the OpenAI model, and the model should be able to invoke the tool using the provided schema.

Actual: The OpenAI API rejects the tool call with a 400 Bad Request error because the schema is invalid. The specific error indicates that additionalProperties is required and must be false in the tool's parameter schema.

Logs

Example OpenAI Error Log:

ERROR    API status error from OpenAI API: Error code: 400 - {'error': {'message': "Invalid schema for function 'add_issue_comment': In context=(), 'additionalProperties' is required to be supplied and 
         to be false.", 'type': 'invalid_request_error', 'param': 'tools[0].function.parameters', 'code': 'invalid_function_parameters'}}                                                                 
WARNING  Attempt 1/1 failed: Invalid schema for function 'add_issue_comment': In context=(), 'additionalProperties' is required to be supplied and to be false.  

Additional Context:

The root cause appears to be the schema generation logic within the underlying mcp-go framework (github.com/mark3labs/mcp-go), which doesn't add the required additionalProperties: false property to the top-level parameter object for most tools defined via mcp.NewTool.

It's worth noting that while tools like PushFiles (repositories.go) and CreatePullRequestReview (pullrequests.go) correctly define additionalProperties: false for the schema of items within their specific array parameters (using mcp.Items), this specific fix doesn't apply to the top-level parameters object of those tools, nor does it apply to the simpler tools (like add_issue_comment) that lack such nested structures. The core issue remains the generation of the main tool parameter schema required by the OpenAI API.

A fix within the mcp-go framework seems necessary for broad compatibility.

[williammartin]

Thanks for the detailed write up.

In my (pretty uninformed) opinion, OpenAI should fix their function calling compatibility to work with the MCP spec which does not require additionalProperties to be set, or to be false.

I'll discuss with the team what we should do here in the meantime.

---

Separately, I wasn't able to reproduce this failure in VSCode:

➜  docker run -i --rm ghcr.io/github/github-mcp-server ./github-mcp-server --version
GitHub MCP Server
Version: v0.2.1
Commit: 9fa582d8d63522d70ce8f3af58265effb9645323
Build Date: 2025-04-21T23:03:01Z

See 4.1 Preview being used.

What obvious thing I am I missing?

[nknj]

OpenAI only requires "additionalProperties": false when using "strict": true (i.e. when using structured outputs). By default the chat completions api sets "strict": false and "additionalProperties": false shouldn't be required.

[williammartin]

@p-breslin can you describe a little bit more how you're actually calling the MCP server? You described:

Configure an MCP client to use this server with an OpenAI model that supports function/tool calling (e.g., gpt-4.1-nano).

But as per #376 (comment), it seems like you're not using chat. It would help to have a little bit more information on your use case here.

[p-breslin]

Yes, #376 (comment) is correct in that I was using structured outputs, apologies for leaving out that detail!

To clarify, I wasn’t using an IDE like VS Code in this case. I encountered the issue while programmatically invoking the MCP server using an agent framework (Agno). I would have thought that this issue shouldn’t depend on whether the request comes from an IDE or another client, as long as the payload and headers are equivalent, the server’s behavior should be consistent. That said, correct me if I’m wrong, perhaps there’s something about how the framework is structuring the request that’s triggering this. Happy to take it up with Agno if so.