The original headline is about an MCP server crashing and an error recovery pattern that fixed it. Let me rewrite this following the rules.

Short summary

A developer shares a battle-tested error recovery pattern for MCP (Model Context Protocol) servers that prevents silent crashes and hallucinated responses. The key insight is wrapping tool handlers with typed exception handling and always setting isError: True in responses so AI clients don't misinterpret error messages as valid output. The pattern covers network failures, bad arguments, and unexpected exceptions with concrete Python code.

• •Wrap MCP tool handlers with typed exception handling to prevent silent server crashes
• •Always set isError: True in error responses or AI clients will parse errors as valid output
• •Reconnect transport on network failures instead of restarting the entire server process

Generated with AI, which can make mistakes.