AI Pulse · Under the Radar

A developer named Teng Li built a linter called mcpgrade that tests how well MCP servers work with live agents. He ran it against 36 popular servers and found something striking: official integrations from MongoDB, Notion, and Airtable all scored D or F grades. The problem is not the underlying APIs. It is the schema documentation agents see when they decide which tool to call.

Most MCP servers auto-generate their schemas from zod validators or OpenAPI specs, and that process strips out the parameter descriptions developers wrote. An agent looks at a tool called create_document with no explanation of what workspace_id means, guesses wrong, and fails. Li measured a 50% out-of-scope task attempt rate on poorly documented servers and 0% on servers with clean, descriptive schemas. Context7 saw the audit, fixed its schema, and jumped from a C to a perfect score in one update.

This matters because MCP is becoming the standard way to connect agents to external tools, and if your server hands an agent a cryptic schema, the agent will misuse it. The fix is straightforward: write clear descriptions for every parameter and result field, test the schema with a real agent before you ship, and do not rely on auto-generation to preserve that context. If you are building or auditing MCP servers at your company, the original report includes a scoring rubric you can use as a checklist. We are folding similar schema audits into the way we build agent workflows in the Omni Command Centre, because a tool an agent cannot understand is a tool it will not use correctly. The gap between a working API and a working agent integration is smaller than most teams think, but it is not zero.