A scan of 3,923 public MCP servers revealed that approximately one in four tools have parameters lacking descriptions, forcing AI models to guess their function. Additionally, 41% of servers (excluding Go) provide raw exception text instead of helpful error messages, hindering model recovery. While direct tool-poisoning attacks were rare, a bug related to Python None defaults affected 8.6% of Python servers, with some fixes already confirmed upstream. AI
IMPACT Highlights critical gaps in tool descriptions and error handling that hinder AI agent performance.
RANK_REASON Analysis of a large dataset of MCP servers using a custom tool. [lever_c_demoted from research: ic=1 ai=0.7]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →