# Robin Saige — the MCP tool observatory > Continuous, reproducible verification of the agent-tool ecosystem. Free, no key. > Method is public and every number is re-derivable. Trust is not requested; recomputability is offered. ## What this is Robin Saige probes every server in the official MCP registry and publishes what actually answers — alive, auth-walled, or dead — with drift tracking and (in the trade lane) truth checks against primary sources. It is itself an MCP server, listed in the registry it measures as com.robinsaige/verifier. ## MCP endpoint https://robinsaige.com/mcp (streamable-http, protocol 2025-06-18, no auth) Tools: - registry_pulse() — the ecosystem at a glance: does it answer + is the answer true (52/52) - find_servers(query) — discover a server by a fragment of its name → get the exact name - check_server(name) — one server's full record incl. whether its answers are verified true - list_findings() — ruled findings (each with a falsifier) + the corrections log ## House rules (the contract) - Grades are strict: observed (we probed it) / reported (registry claims it) / derived (we computed it). - A "true" verdict requires a primary source; otherwise at most "consistent" or "stable". - Auth-walled servers are alive but unverifiable from outside — never scored dead, never verified. - A finding carries a falsification test or it is not a finding. - Corrections are public and never deleted. The 27%->55% correction is on /corrections. - We identify ourselves when probing and never bypass authentication. ## Pages / overview · /truth Tier-2 truth checks vs primary sources · /atlas ruled findings · /signals the traffic tripwire · /find look up any server · /data (every number + its SQL) · /method · /corrections ## What we found (2026-08 census) 55% of remote servers answer a real handshake; a quarter demand a login; one gateway wears 1,312 listings. 52 of 52 answers we could re-derive against a public primary source (USITC tariff schedule, Eurostat Comext, ICC/WCO/UNECE) matched — one to the euro.