Remote MCP Server | Session Terminated after 1st interaction

[hayescode]

Describe the bug

I'm using OpenAI Responses API and the remote GitHub MCP server with PAT authentication. Starting a few days ago I noticed that the first interaction works fine, but then every subsequent tool call fails. This used to work fine.

Some users report setting stateless_http=True in the MCP server itself seems to fix this issue (Source).

    "name": "list_pull_requests",
    "server_label": "github",
    "type": "mcp_call",
    "error": {
      "type": "mcp_protocol_error",
      "code": 32600,
      "message": "Session terminated"
    },

Affected version

latest

Steps to reproduce the behavior

response = await llm.responses.create(
    model="gpt-4.1",
    tools=[
        Mcp(
            server_label="github",
            server_url="https://api.githubcopilot.com/mcp/",
            type="mcp",
            headers={
                "Authorization": f"Bearer {GITHUB_PAT}",
            },
            require_approval="never",
        )
    ],
    input="what is my github username? how many issues do i have open?"
)
print(response.model_dump_json(indent=2))

Expected vs actual behavior

The session should persist for multiple tool calls.

[JhandryChimbo]

The GitHub MCP server is likely stateful by default, meaning it tries to maintain a session across tool calls. If the session gets dropped or expires (which can happen between requests), you'll see the Session terminated error.

Fix

You should explicitly set the stateless_http=True parameter when defining the MCP tool. This tells the MCP server to treat each request independently, avoiding session issues.

Here’s the fixed version of your code:

response = await llm.responses.create(
    model="gpt-4.1",
    tools=[
        Mcp(
            server_label="github",
            server_url="https://api.githubcopilot.com/mcp/",
            type="mcp",
            headers={
                "Authorization": f"Bearer {GITHUB_PAT}",
            },
            require_approval="never",
            stateless_http=True  # This is the fix
        )
    ],
    input="what is my github username? how many issues do i have open?"
)
print(response.model_dump_json(indent=2))

[hayescode]

@JhandryChimbo stateless_http is not an option see https://github.com/openai/openai-python/blob/main/src/openai/types/responses/tool_param.py#L61-L84

[JhandryChimbo]

Sorry, I thought I used it before.

[hayescode]

bump. still getting this issue on the latest OpenAI python sdk 1.99.1

[LuluBeatson]

Hi @hayescode, thank you for raising this issue

The Remote GitHub MCP server does use sessions. i.e. the client must initialize a session, getting back a session_id and use a Mcp-Session-Id header for tools/list and subsequent tool calls. The MCP specification describes this "Session Management"

If the core OpenAI python library does not support MCP sessions, you could also try the OpenAI Agents SDK. You could try the MCPServerStreamableHttp (example) but I also see an open issue related to MCP session support: openai/openai-agents-python#924

Please let us know if you're able to resolve this, and we will also keep an eye on MCP session support in the OpenAI libraries.

[hayescode]

Thank you very much @LuluBeatson ! This is great information and it's all making sense now. Thank you!

[mattdholloway]

@hayescode Are you happy for this issue to be closed?

[hayescode]

@hayescode Are you happy for this issue to be closed?

@mattdholloway I would prefer session management to be settable in the headers similar to X-MCP-Readonly Boolean header. While the root cause makes sense and the bug seems to be in OpenAIs implementation, I don't see the benefit of requiring sessions and my other authenticated MCP servers work fine without it. Would that be possible?