Every remote server in the official MCP registry, probed live and graded with receipts — the only MCP index where every grade shows its evidence.
1 match. Categories are keyword-derived from registry metadata — imperfect by design, deterministic by principle.
| # | Server | Grade | Score | Tools |
|---|---|---|---|---|
| 1 | ai.openhelm/deep-research | D prov. | 52 | — |
Each server is probed live over streamable HTTP: reachability (25) — does initialize succeed; auth-gated servers earn partial credit only if they advertise WWW-Authenticate so clients can discover OAuth. protocol (15) — valid JSON-RPC handshake; deprecated SSE transport is penalized. tooling (35) — tools/list works, share of tools with descriptions and fully-typed input schemas, median description depth. latency (10) — initialize round-trip. provenance (15) — registry metadata completeness and whether the reverse-DNS namespace actually matches the serving domain. Scores scale to what is verifiable; unverifiable dimensions mark the grade provisional rather than guessing. Every point carries the verbatim observation that earned it. No stars, no votes, no pay-to-rank — probes only.