MCP Liveness

A registry listing is a claim the publisher typed. This is whether it answers.

curl "https://agentwares-mcp-liveness.vercel.app/v1/alive?name=io.github.owner/server"
curl "https://agentwares-mcp-liveness.vercel.app/v1/alive?url=https://example.com/mcp"   # not in the registry? check the endpoint

One handshake — server/discover, then initialize for a server on the older revision — no credential, no retry. Six answers: usable, needs_credential, charged, unreachable, local_only, not_listed — each with a one-line fix. A check by URL answers the first four; the URL must be public https, and private, loopback, link-local and cloud-metadata addresses are refused, redirects included.

Every remote listing in the registry also has its own status page at /servers/<registry name> — a live check, its monthly census result and what changed in its tools, for example io.github.agentwares/mcp-liveness.

Email me when this server changes

A check is one look, now. To hear when a server stops answering or its tools change, leave an address: agentcheck checks it every hour, for free, with no password and no GitHub.

Free, no password or GitHub: one confirmation email first, then a check every hour and an email when it stops answering or its tools change. Every email has a link to stop. More about it.

Agents can do the same with agentcheck_create_target on agentcheck’s MCP server, or POST https://agentwares-agentcheck.vercel.app/api/v1/watch with a server and an address.

Publish the server yourself? mcpcheck runs scheduled checks on it from outside: clean-install and tool-catalog checks weekly on Server ($19 a server a month), or nightly with the client matrix and OAuth conformance added on Server Pro ($49 a server a month). The first run is free.

What a 401 is hiding

Two servers can both answer 401 and be nothing alike. One publishes the discovery chain the spec defines, so a client nobody has heard of can register itself and a person approves once. The other wants a key from a dashboard. The registry lists them identically. On a needs_credential answer this follows RFC 9728 → RFC 8414 → RFC 7591 and says which one you are looking at.

Your own server’s sign-in failing? The sign-in check walks the whole chain a client walks and names the first broken step, whose it is — the server, its identity provider, or a known bug in claude.ai, Claude Code, ChatGPT, VS Code or Cursor — and the fix. Error messages explained: “Authorization with the MCP server failed” · “Couldn't reach the MCP server” · “Failed to discover OAuth metadata” · “Dynamic Client Registration not supported”.

Your MCP App renders blank? The MCP Apps check reads each tool’s ui:// view the way Claude, ChatGPT and VS Code do, names the first step that stops it rendering and whose it is, checks what ChatGPT’s directory review checks, and shows the view in a sandboxed preview. Error messages explained: “MCP App renders blank” · “Resource not found: ui://” · “Tool has no UI” · “Invalid ui.domain format”.

The population, 2026-10-06

Of the 263 that want a credential, 204 publish the full chain and a registration endpoint, 11 need a person to create the client first, and 45 published nothing a stock client could discover at all.

A random sample of 1200 of 25,212 remote listings, out of 40,105 distinct servers in the registry. Full series: /v1/survey.

Do not probe the catalogue yourself

Every remote listing is probed once a month and the answers are served as static files. Reading them is one request instead of nineteen thousand, and it costs the servers being asked about nothing at all.

curl "https://agentwares-mcp-liveness.vercel.app/unreachable.ndjson"   # only the ones to skip
curl "https://agentwares-mcp-liveness.vercel.app/index.ndjson"         # all of them
curl "https://agentwares-mcp-liveness.vercel.app/index.meta.json"      # counts, causes, when it was built

/v1/alive probes live and is for one server you care about right now. The files are for choosing between many.

Corrected 8 October 2026. The census used to ask only initialize, so servers that speak only the 2026-07-28 revision of the spec were counted unreachable. Every check now asks server/discover at 2026-07-28 first and initialize second, and initialize asks 2025-11-25 rather than 2025-06-18. The 10 listings whose 6 October reason was a version refusal were re-probed: 7 are usable (6 on 2026-07-28, 1 on initialize) and 3 refuse both. So were the 147 that answered a bare HTTP 400, which is how a server that speaks only 2026-07-28 refuses initialize: 64 are usable (41 on 2026-07-28, 23 on initialize at 2025-11-25), 2 want a credential and 81 stay unreachable. Both sets are named under corrections in index.meta.json, which also counts usable listings by the handshake that answered, in usableByEra.

Agent discovery documents. Each census also asks every origin behind the listings for an A2A agent card, an AI Catalog and the MCP server cards it lists (how SEP-2127 finds them), an MCP server card (at the location SEP-2127 reserves, and where the superseded SEP-1649 draft put it) and a Web Bot Auth key directory, one GET each with no credential, and counts who publishes them and who validates: who publishes agent cards, server cards and bot-auth keys.

Cross-checked. On the 2 October 2026 census, 23 of 23 verdicts agreed with Agent Nexus’s independent probes — every endpoint both of us measured between 1 and 3 October. A clean 401 or 403 counts as up on both sides.

What it will not do

It never sends a credential, never registers a client, never retries a refusal and never sends another operator’s user-agent. Every request says agentwares-mcp-liveness/0.1 (+https://agentwares-agentcheck.vercel.app/bot). Given a URL, it never fetches inside a private network: the host is resolved and every address and every redirect is checked first, and one caller gets 30 URL checks a minute per server instance.

Also

MCP server · llms.txt · server.json · survey

Free, no key, no signup. Built by agentcheck, which monitors servers you own.