A client connected to the hosted MCP server at api.ax1om.ai/mcp was cut off
every five minutes and reconnected in a loop, which made the surface unusable
for any session longer than a short exchange. The cause was the platform’s
request timeout meeting a long-lived stream the server held open.
The server has always run stateless, which is the right shape for a service that autoscales with no session affinity: every tool call is a request and a response, grants are resolved per request, and consent lives in the database. That mode needs no long-lived stream, so the server no longer holds one open past the platform’s limit, and a session is as long as the conversation. Three smaller things that made the failure hard to see in the logs were fixed with it.