3 MCP server failure modes that bit us in production, and how we ship around them

TL;DR AI
2 min readKey summary
A postmortem on MCP production issues describes three recurring failure modes in real deployments.
Built-in agent tools could silently shadow sandboxed MCP tools, creating security and reliability risk.
Environment variables could leak into child MCP processes, raising the chance of token exposure.
The team now uses explicit allowlists and disallowlists to keep agents in the intended workspace and protect sensitive data.
