一名开发者利用自主创建的mcpgrade评分工具，对36个流行的MCP（模型上下文协议）服务器进行了AI代理可用性测试 [1]。结果显示，15个服务器获得A评分，而11个服务器获得D或F评分 [1]。
评测中暴露的问题主要并非协议合规性缺陷，而是集中在参数文档不足、工具命名不规范和模式架构设计不当等方面 [1]。这些缺陷导致AI模型在工具选择、参数填充以至工具使用上存在明显偏差。
具体而言，MongoDB官方服务器检出66个错误，Notion官方服务器62个错误，Airtable 69个错误，todoist-mcp-server 110个错误，firecrawl-mcp则最低，评分仅57分且含134个错误 [1]。其中参数描述缺失（D004规则）是最普遍的问题，firecrawl服务器在此项就有132个错误，todoist则有110个 [1]。
文档质量对AI模型表现的影响显著。在文档详尽的服务器上，工具选择准确率达到100%，但在firecrawl上则降至84% [1]。当目录不够明确时，模型正确拒绝范围外任务的准确率仅为50%；而在文档清晰的小型目录上，该准确率可达100% [1]。
shrimp-task-manager的表现证明了即使是包含15个工具的庞大目录，通过严格的文档维护也能获得A评分（96分），说明规模本身并非决定性因素 [1]。
A developer has evaluated the usability of 36 Model Context Protocol (MCP) servers by creating the mcpgrade assessment tool, uncovering significant quality gaps in how well AI agents can interact with them. [1] The evaluation revealed that 15 servers earned A grades while 11 received D or F ratings, with problems concentrated in parameter documentation, naming conventions, and schema design rather than protocol compliance failures. [1]
Poor documentation quality directly undermines agent performance. On servers with detailed documentation, the model's tool selection accuracy reached 100%, but this dropped to 84% when testing against firecrawl-mcp, which received the lowest score at 57 points. [1] The same documentation effect appears in task rejection: models correctly declined out-of-scope requests 100% of the time on servers with clear, small tool directories, but only 50% of the time on those with ambiguous, larger catalogs. [1]
The most common failure mode stems from inadequate parameter descriptions. MongoDB's official server contained 66 errors, Notion's 62, and firecrawl recorded the highest count at 134, with 132 of those specifically due to missing parameter documentation. [1] Todoist-mcp-server reported 110 such errors, while Airtable recorded 69 total errors. [1] These gaps cause AI models to select inappropriate tools, mishandle parameters, or skip tools entirely.
Notably, the shrimp-task-manager server demonstrates that large tool directories can maintain high usability through rigorous documentation standards, achieving an A grade with 96 points across 15 tools. [1] This suggests the quality problem is solvable rather than inherent to larger server implementations.