We read the schemas of 4,951 public MCP servers

说明与用途

研究发现，21.8% 的参数缺少描述，17.4% 的工具描述相同，且模式大小可达 280,000 个 token。提供了改进描述和检查冲突的建议。

问题

MCP server tool schemas often lack descriptive clarity, leading to ambiguous tool selection and potential model errors.

结果

The study revealed widespread schema deficiencies, quantified their prevalence, and offered concrete guidance for improving tool and parameter descriptions to reduce model ambiguity.

工作流

• Collect tool schemas from Smithery registry or local servers
• Validate registry data against real server responses
• Analyze parameter descriptions for presence and distinctiveness
• Compute tool description distinctiveness per server
• Measure token cost of schema transmission
• Identify enum usage and enforceability
• Generate actionable recommendations for server maintainers

证据

59,038 of 270,487 parameters — 21.8% — ship with no description.

来源

17.4% of tools are exactly zero distinctive.

来源

Median server: 4,991 bytes, roughly 1,250 tokens

来源