Does the new version responseapi support the mcpserver which is deployed by myself?

[Rama-19]

Does the new version responseapi support the mcpserver which is deployed by myself？Why I change the server_url to my mcpserver's url , it did not work ?

[manimovassagh]

The Responses API does support custom/self-hosted MCP servers, but there are a few requirements that commonly trip people up:

1. Your server must implement the Streamable HTTP transport

OpenAI's servers connect to your MCP endpoint directly. They expect the MCP Streamable HTTP transport (not stdio, not the older SSE-only transport). Your endpoint should handle POST requests with JSON-RPC payloads and respond accordingly.

If you're using FastMCP, make sure you're running it with the HTTP transport:

from fastmcp import FastMCP

mcp = FastMCP("my-server")

@mcp.tool()
def my_tool(query: str) -> str:
    return f"Result for: {query}"

# Run with HTTP transport
mcp.run(transport="http", host="0.0.0.0", port=8000)

2. The server must be publicly reachable from OpenAI's infrastructure

When you pass server_url to the Responses API, it's OpenAI's servers (not your local machine) that connect to that URL. localhost or private IPs won't work. You need a publicly accessible URL.

Quick options for testing:

• Use ngrok: ngrok http 8000
• Deploy to a cloud VM with a public IP
• Use a service like Railway or Fly.io

3. HTTPS is strongly recommended

Some MCP configurations require HTTPS. Use a reverse proxy with TLS or a tunnel service that provides it.

4. Verify your endpoint externally

Test from a machine that isn't your server:

curl -X POST https://your-public-url/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc": "2.0", "method": "tools/list", "id": 1}'

If this returns your tool list, OpenAI should be able to reach it too. If it doesn't, the issue is network/transport configuration, not the SDK.