MCP Server Session Terminates After 1st Interaction

[hayescode]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

When using authenticated remote MCP servers like GitHub's remote MCP the list_tools() runs successfully however actually calling the tools produces the following error.

{
    "id": "mcp_68938c85a9908192bfa5ffe1c83af8f40446391e66c97710",
    "arguments": "{\"owner\":\"<REDACTED>}",
    "name": "list_pull_requests",
    "server_label": "github",
    "type": "mcp_call",
    "error": {
        "type": "mcp_protocol_error",
        "code": 32600,
        "message": "Session terminated"
    },
    "output": null,
    "approval_request_id": null
},

If I use the previous_response_id that contains the MCP list_tools then the 1st tool call works, but the second gives the same session terminated error.

So it seems like the MCP server is getting disconnected and the session is not persisted. This only occurs for authenticated remote MCP servers.

References:

• https://community.openai.com/t/error-using-mcp-server-using-open-ai-api-playground/1280337/2
• https://community.openai.com/t/the-responses-api-terminates-a-session-too-early/1312539
• https://community.openai.com/t/openai-responses-api-to-mcp-server-connection-problems/1269983
• Remote MCP Server | Session Terminated after 1st interaction github/github-mcp-server#702
• FastMCP does not work with OpenAI's responses API jlowin/fastmcp#855

To Reproduce

from openai.types.responses.tool_param import Mcp

mcp_github = Mcp(
    server_label="github",
    server_url="https://api.githubcopilot.com/mcp/",
    type="mcp",
    headers={
        "Authorization": f"Bearer {GITHUB_PAT}",
    },
    require_approval="never",
)

stream = await llm.responses.create(
    model="gpt-4.1",
    tools=[mcp_github],
    input="What is my GitHub username?",
    stream=True
)
async for event in stream:
    print(event.model_dump_json(indent=2))

Code snippets

OS

Windows 11

Python version

Python v3.13.2

Library version

openai v.1.99.1

[harjothkhara]

Thanks for raising this! The problem lies in the SDK closing the HTTP connection after listing tools. When an MCP tool is used, the session on the remote server is tied to that connection. Closing it causes the server to report mcp_protocol_error 32600 ("Session terminated"). This is why only the first interaction works.

I've opened PR #2526 that fixes this by skipping the stream's .close() call when an MCP tool is present. The change ensures the session persists across multiple tool calls, addressing the issue described here and in the related GitHub MCP server issue (#702).

[hayescode]

https://community.openai.com/t/the-responses-api-terminates-a-session-too-early/1312539