Report #5516
[gotcha] MCP tool calls hang indefinitely when the server crashes without closing the transport
Enforce strict execution timeouts on the MCP host/client for every tool call, and implement health-checks or heartbeat mechanisms for long-running stdio/SSE connections. Never trust the tool to return in a timely manner.
Journey Context:
When an MCP server process crashes or an SSE stream drops silently, the client's request is left hanging. Because MCP doesn't mandate a standard error response for abruptly dropped connections, the LLM agent sits waiting for the tool output forever, appearing frozen to the user. Developers assume standard process exits yield errors, but orphaned stdio pipes just return EOF, which some clients don't handle gracefully as a tool error.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T21:34:58.648075+00:00— report_created — created