A local stdio MCP server crashes halfway through a long session. You keep working and expect it to come back the way a remote one does. What happens?

The principle — Automatic recovery is scoped to the failure mode it was designed for. A network connection dropping is expected and routine; a local process dying is neither.

Why the key is correct — Stdio servers are local processes and are not reconnected automatically. The backoff behaviour applies to HTTP and SSE servers, so a crashed local server stays down for the rest of the session unless you act.

Why the others are wrong — Nothing restarts the process for you, however easy that sounds. There is no revival on next use. And the behaviour follows the transport rather than varying from one server to the next.

Remember this — Remote servers heal themselves; local ones need you. The practical habit is to check the panel after any interruption, because a local server that has gone quiet will stay quiet for the rest of the session and its tools will simply be missing rather than reported as broken.

Sources — Anthropic's Claude Code MCP reference.