Issue with custom MCP server

[ksawaray]

Hi there,

I'm encountering an issue when trying to use a custom MCP server created with fastMCP for tool calls via client.responses.create().

• OpenAI version: 1.82.1
• Python version: 3.11.10

from openai import OpenAI

openai_api_key = os.getenv("OPENAI_API_KEY")
client = OpenAI(api_key=openai_api_key)

messages = [{"role": "user", "content": "xxxxxxx"}]

tools = [
        {
            "type": "mcp",
            "server_label": "<server label>",
            "server_url": "http://<my domain>:<port>/mcp",
            "require_approval": "never",
        }
]

response = client.responses.create(
        model="gpt-4o",
        input=messages,
        tools=tools
)

When I execute the code, I get the following error.

openai.APIStatusError: Error code: 424 - {'error': {'message': "Error retrieving tool list from MCP server: 'xxxxxxx'. Http status code: 424 (Failed Dependency)", 'type': 'external_connector_error', 'param': 'tools', 'code': 'http_error'}}

• I confirmed that the URL http://:/mcp is publicly accessible.
• When I changed the tool type to "deepwiki" specified in the official documentation (https://platform.openai.com/docs/guides/tools-remote-mcp), it worked correctly.

Could you please help me understand the root cause of this 424 Failed Dependency error when using a custom MCP server, and how to resolve it?
(Is the issue perhaps related to this library sending requests to https://api.openai.com/v1/responses?)

[Toygarmetu]

The 424 error means OpenAI's servers are trying to reach your MCP server from their network and failing — not from your machine.

When you pass an MCP server URL to client.responses.create(), OpenAI's backend makes the connection to your MCP server directly (to fetch the tool list and invoke tools). Your local machine is just sending the configuration to the API.

Requirements for a self-hosted MCP server to work:

1. Publicly reachable URL — The server must be accessible from OpenAI's infrastructure, not just from your local network or behind a NAT/firewall. http://localhost:... will never work. A public IP or domain with port forwarding is required.

2. Correct MCP transport protocol — OpenAI expects the HTTP+SSE transport (Server-Sent Events). Your /mcp endpoint must implement the MCP spec's SSE transport, not a custom HTTP API.

3. HTTPS recommended — While not always enforced, some OpenAI integrations require HTTPS. Use a reverse proxy (nginx + Let's Encrypt) if needed.

Quick check:

curl -N -H "Accept: text/event-stream" https://your-domain.com/mcp

If this works from an external machine (not your server), OpenAI should be able to reach it too.

Debugging: The error message includes your server label ('xxxxxxx'). The 424 specifically means OpenAI reached your endpoint but got an unexpected HTTP status code back — check your MCP server's response to the initial tool-list request.