How we rank public MCP servers from the outside.
One score, 0 to 100. Did we reach the server, call a tool, and get a result that looks like the job, and is that faster than the other servers on the Index. See the live scores.
Four signals, then a host gate.
Letters: A is 90 or above, B 80, C 70, D 60. Below 60 is F. Inputs are the last 7 days of probes. Public monitors commonly run about every 6 hours.
score = health × (0.5 + 0.5×host pass)
If no named host was measured, the host term is 1 and the score is the health mix. If tool success is missing, drop its 30% and divide the rest by 0.70.
Reachable runs divided by total runs over the last 7 days.
Invocation × 0.6 + functional × 0.4. A completed call, and a result that looks like the job.
Share of scored servers with a slower P95 than this one. Relative to the Index, not a fixed time lid.
After we negotiate, did transport, discovery, and invocation look like MCP.
Host pass is the average across hosts we actually ran. A full pass is 100, a partial walk is that host's tool rate, a hard fail is 0. Untested hosts are left out. If every measured host fails, the score is half the health mix, so it cannot be an A.
What each column answers.
Reach (7d)
What share of runs reached the server?
Tool success
Did the call complete, and did the result look like the job?
Speed
How does this P95 compare with the other scored servers?
Spec
Did the path look like MCP after we negotiated?
Who gets a rank.
Ranked
Too new
No score
Popular
Identity
Protocol era and tool class are labels on the probe, not extra weights. The per-host record is Host Compatibility.
What we are, and are not.
We are not
- • A paid placement directory
- • A full MCP spec certification
- • A ranking of Cursor vs Claude vs VS Code
- • An inside-out APM product
We are
- • An outside-in check of public MCP URLs
- • One published score with four weights
- • A host gate when we have measured those hosts
- • Honest that listed is not the same as called
Common questions.
| Question | Answer |
|---|---|
| Is this a lab benchmark of MCP servers? | No. We probe production MCP URLs on a schedule and score whether the path held: reach, list tools, call a safe tool, and whether the result looked like the job. |
| What goes into the Score? | Four signals: reach 35%, tool success 30%, speed 20%, and spec 15%. If tool success is missing, the other three are rescaled. If named hosts were measured, that mix is then multiplied by (0.5 + 0.5 times host pass rate). Untested hosts do not count. |
| How is speed scored? | Speed is how many other scored servers this one beats on P95 latency. It is a rank inside the current Index, not a fixed millisecond lid. |
| What does tool success mean? | Did the call complete, and did the result look like the job? Those two rates are blended 60% and 40%, then that blend is 30% of the Score. The default path uses safe read tools. |
| Do you rank Cursor vs Claude vs VS Code? | No. Hosts are a gate on the server's score, not a leaderboard of hosts. The per-host walk is Host Compatibility. |
| Why is a server not ranked? | Fewer than 2 probe runs in 7 days shows a score without a rank. Login-required and not-rated servers get no score. |
| Can a server pay for a better rank? | No. Rankings come from measured results only. Paused monitors are hidden from the public index. |
| What is Popular? | A filter, not part of Score. It adds Smithery uses, PulseMCP visitors, npm last-month downloads, PyPI last-month downloads, and Docker Hub pulls. Registry and Glama are listings only. |
Trust is measured, not purchased.
The Public MCP Score is operated by Carmel Labs / AgentStatus. Rank comes from the probes. The only way to move up is to improve reach, tool success, speed, and spec under the same outside-in path.
