I checked 22,561 MCP servers. Almost none have a reliability record. Here's how to vet one before you ship.
A recent analysis of over 22,500 Multi-Call Protocol (MCP) servers revealed a significant lack of reliability data, with only about 0.5% having any independent runtime observation. Many of these servers, even those from reputable companies, score poorly on performance metrics like latency and success rates. The author suggests that popularity metrics like GitHub stars are insufficient for vetting these dependencies, and recommends a practical checklist including direct testing, checking recency, and treating tool descriptions with caution. A proposed solution involves using an independent trust score before an agent makes a tool call to mitigate risks associated with unreliable MCP servers. AI
IMPACT Highlights critical infrastructure risks for AI agents relying on external tools, necessitating new vetting processes.