Outside-in rankings for public MCP servers.
We score public MCP URLs from the outside: reach the server, negotiate the dialect, list tools, complete a call, and record whether the result looked like the job. Not a URL ping. See methodology for weights, and why Displayed Score can differ from Rank Score.
Updated Sep 9, 2026·158,535 runs·3289 servers·~every 6 hours
What each probe does
01
Negotiate
Modern 2026-07-28 path first (server/discover + per-request meta); legacy initialize/session if required.
02
Discover
tools/list (+ resources & prompts when the server exposes them).
03
Call
Safe tools/call — up to 20 read-only tools, plus schema-linked workflows.
04
Score
Invoke (accepted & well-formed) vs functional (response quality).
05
Composite
30d reach · P95 · tool blend · spec → Score. Same suite for every server.
Default index path exercises safe reads with synthesized args — not your full docs/contract suite. Deeper contract probes are available on opted-in agents.
Protocol path filters by which MCP dialect succeeded on the latest probe, not by health color. Status shows two chips: colored health (Operational / Impaired / Down / Auth) and a neutral protocol chip (Modern, Legacy, Dual, Modern only, …).
Overall Score
Five signals from the protocol path, not uptime alone. The Score column is Rank Score (confidence-adjusted for low sample). Click a row for Displayed Score, anatomy, and per-tool evidence.
