I’m experiencing an intermittent issue while connecting my custom MCP server to ChatGPT.
Expected behavior
ChatGPT successfully discovers the OAuth configuration and returns the OAuth metadata (authorization URL, token URL, registration URL, PKCE configuration, etc.), after which the OAuth flow completes successfully.
Actual behavior
Occasionally, the same request fails with:
{
"detail": "MCP server https://mcp.birdeye.com/mcp does not implement OAuth"
}
Observations
- The issue is intermittent. Retrying the exact same connection sometimes succeeds without any changes.
- When the request succeeds, ChatGPT’s
POST /backend-api/aip/connectors/mcp/oauth_configendpoint returns the expected OAuth configuration. - When it fails, the same endpoint returns the “does not implement OAuth” error.
- Most importantly, during failed attempts, no HTTP request reaches our MCP server (including the MCP endpoint, OAuth discovery endpoints, or any OAuth endpoints). Our server access logs show no incoming requests at all.
- The exact same MCP server and OAuth implementation (Dynamic Client Registration + PKCE) works consistently with Claude and other MCP clients.
This makes it appear that the failure is occurring within ChatGPT’s OAuth discovery or connector backend before any request is sent to the MCP server.
Has anyone else experienced this intermittent behavior, or can the OpenAI team confirm whether this could be related to OAuth discovery caching or the MCP connector infrastructure?
I’ve attached screenshot of failed response.
In case of failure, this api return 404 with response body: {“detail”:“MCP server [my mcp server url here] does not implement OAuth”}. In this case there no API call at my MCP server.
