Is OpenAI’s support (platform, responses API, etc.) for MCP considered to be in beta, or is it final? I ask because I have encountered numerous issues that render it virtually unusable:
- The behavior of the integrated client differs when I use the responses API compared to the playground. The responses API merely sends a POST request to the MCP endpoint, whereas the playground attempts GET requests. If, as per protocol, the server responds with a 405 (method not allowed), the playground returns an error and breaks.
- The client does not appear to support progress tokens; it neither passes this parameter to the server nor acknowledges keep-alive notifications. This makes it entirely impossible to use tools for lengthy agent operations, resulting in a timeout error—even though the server could appropriately report a progress status.
- There is also no way to set a timeout for requests, which might have served as a workaround for the previous issue.
- Documentation is incomplete and lacking details about the client behavior and supported features.
For these reasons, I have found OpenAI’s MCP implementation to be practically unusable for my purposes.