I lint-scanned 36 popular MCP servers. A third of them are failing your agent.
Summary
I lint-scanned 36 popular MCP servers to evaluate their usability for AI agents. The study found that many servers pass protocol specs but fail on descriptions and documentation, with undocumented parameters being a major source of errors, and demonstrates how tooling design affects agent performance. The article provides a Lighthouse-style mcpgrade scorecard, examples of top and bottom performers, and practical guidance for improving tool descriptions and schemas.