I graded 36 popular MCP servers on agent usability. A third got a D or F

A developer created "mcpgrade," a scorecard tool for evaluating Model Context Protocol (MCP) servers' usability by AI agents, and found that one-third of 36 popular servers received failing grades despite being spec-compliant. The primary issue across failing servers is missing parameter descriptions in their schemas—developers use code generators like Zod or OpenAPI but fail to add descriptions, leaving AI models without crucial information needed to correctly select and use tools.