OpenAI fails to load tools

[m4recek]

I have a deployed MCP server on public k8s, it can integrate with Cursor through mcp-remote, it successfully loads tools, and I can execute operations on the GraphQL schema. Auth is handled through Auth0.

But when I add it to OpenAI with an Access token / API key, which is valid, it just says there are no tools available.

How should I debug this?

[alocay]

Is it able to load the built-in search, execute, validate, and introspect tools but not run-time generated tools from GQL operations? Or is not loading even the built-in tools?

Also can you provide a bit more details on your server configuration?

[m4recek]

I updated our deployment to 1.0 yesterday and kinda started working....

• It can load built-in tools and can make valid requests when adding MCP to Chat now
• It cannot load built-in tools when I add MCP to Agent-builder - It gets 401 - Unauthorized (with the same key as chat)

.

In Agent builder:

.
.

In Chat:

[m4recek]

Aaaaand it works. I was able to load tools in all "GUIs" of OpenAI now, so it looks like they fixed something, our deployment is still running 1.0 without changes. compared to yesterday.

[m4recek]

It is still intermitent, Open AI is getting 401 every few request, not sure what they are sending but it does not work like 60% of the time.

[alocay]

Are you able to test this with a locally running MCP Server communicating with OpenAI client? Might be easier to debug this way if possible. Also is there anything in the log output regarding the auth error (the log can be routed to a file if needed).

[DaleSeo]

Hi @m4recek, it seems you are experiencing a session affinity issue with your Kubernetes load balancer.

Apollo MCP Server uses a stateful protocol requiring sticky sessions via the mcp-session-id header. Most cloud load balancers don't support header-based session affinity, so OpenAI's requests get routed to different backend instances, breaking the session.

Could you please try enabling stateless mode in your configuration as follows? This will make each request independent and eliminate the need for session affinity.

transport:
  type: streamable_http
  stateful_mode: false

This issue is documented in our deployment docs: https://www.apollographql.com/docs/apollo-mcp-server/deploy#production-considerations. Let us know if enabling stateless mode resolves the issue!