MCP Scoreboard

Choosing an MCP server currently means reading its README and hoping. This crawls the ecosystem — Glama, Smithery, PulseMCP, GitHub topics, Docker Hub, npm, PyPI and the awesome-lists — then scores what it finds across six dimensions: schema quality, protocol compliance, reliability, documentation, security and agent usability.

Scoring is three different kinds of evidence, not one: static analysis of the source, live protocol probes against remote endpoints for latency and uptime, and functional testing inside an isolated Docker sandbox. Publishers get a grade they did not ask for, which is the only kind worth anything. The scoring engine is a separate open-source package, so the grade is auditable rather than a number we assert.