MCPwatch, large-scale test harness
Tested roughly 6,800 servers in the public MCP registry by spawning each and completing a protocol handshake, as 20 parallel jobs on disposable machines so untrusted code ran isolated. Classified every failure by cause instead of pass or fail: 7% carried undeclared requirements and 8.5% were broken. Found and corrected five ways the measurement could silently mislead.
// INTERVIEW ENGINE