Every day a small job on WITAN takes the next slice of the official Model Context Protocol registry and asks each remote server one question: will you complete an MCP handshake? The answers go into a public dataset, MCP server liveness, one record per probe. This post reads one week of it.
Status, plainly: WITAN is a testnet preview built by one developer. These numbers come from one vantage point in South Korea, over one week, and cover the part of the registry this week's slices reached. They say nothing about servers we did not probe.
What we measured
From November 10 to 17, 2026 the job sent 6,400 probes to 4,221 distinct remote endpoints. About 500 endpoints a day are new to the lap; up to 300 more are endpoints whose last probe failed, asked again until they answer. Counting every probe would over-count failures, so the numbers below use each endpoint's latest probe that week.
A probe is the handshake a client makes: initialize (protocol version 2025-11-25), notifications/initialized, then tools/list. No tool is called and no credential is sent. One attempt, 10 seconds per step.
The numbers
| Latest result per endpoint | Endpoints | Share |
|---|---|---|
| Completed the handshake | 2,309 | 54.7% |
| Asked for authorization (401, 403 or a challenge) | 1,294 | 30.7% |
| Answered, but not with MCP | 440 | 10.4% |
| Could not be reached | 178 | 4.2% |
An endpoint that asks for authorization may be perfectly healthy: it just will not talk to a stranger. We stop there and do not guess.
Answered, but not with MCP. 348 of the 440 answered with an HTTP 4xx other than an auth challenge, and 284 of those were a plain 404: the URL in the registry no longer has a server behind it. 46 answered with a 5xx. 29 redirected somewhere we do not follow.
21 endpoints answered the handshake with 402 Payment Required. That is the status x402 uses to ask for payment, before a single tool is listed.
Could not be reached. 136 of the 178 were DNS failures: the host name in the registry no longer resolves. 23 timed out and 13 failed TLS.
The servers that answered
- Protocol version: 1,757 of the 2,309 (76%) answered with 2025-11-25, 405 with 2025-06-18, 80 with 2024-11-05 and 62 with 2025-03-26.
- Tools: the median server listed 7 tools. 69 listed 50 or more; 2 listed none.
- Time to answer
initialize: median 585 ms, 90th percentile 1.3 s, from Seoul. Latency mostly measures the distance to us, not the server.
What this does not show
- One vantage. A server that is unreachable from here may be fine from elsewhere, and latency is shaped by our network.
- One week, one slice. A full lap of the registry takes weeks, so next week's slice is a different set of servers. We do not compare week to week.
- Healthy is not useful. A completed handshake says the server speaks MCP, not that its tools work.
Read it yourself
The dataset is free to read, with or without an agent key. Each version holds every probe so far; filter on measured_at for the latest lap. Over MCP, read_dataset and query_dataset read it, and on the command line:
pip install witan-sdk
wtn query mcp-server-liveness "SELECT error_class, count(*) FROM records GROUP BY 1 ORDER BY 2 DESC" --remote
If you run one of these servers and think a record is wrong, tell us at [email protected] and we will look at it.