App Finishers

Report · 19 November 2026 · published monthly

The state of MCP servers, November 2026

We connected to every company MCP server in the official MCP Registry the way an AI app does, twice, at least ten minutes apart. Most answered. Some never did.

1 in 20company MCP servers in the official registry failed both of our checks.

21,180company servers checked
1,041failed both times
4.9%of the servers we checked

What broke

For each broken server, the answer it gave the second time. Most of these are not subtle bugs. The address itself is gone, or nothing is listening behind it.

Why the 1,041 servers failed, by the answer they gave the second time
CauseServers
The address no longer exists (DNS)27126% of the broken servers
The server crashed or is unreachable (5xx)25524% of the broken servers
Nothing at the listed path (404)24624% of the broken servers
No answer within 10 seconds10710% of the broken servers
A website answers, not MCP (405, HTML)808% of the broken servers
Certificate or TLS error747% of the broken servers
Refuses a standard client (other 4xx)71% of the broken servers
Other1<1% of the broken servers

What each one means, and the fix

The rest (8) failed in other ways, such as asking for a session ID on the very first request.

Servers that only speak the newest protocol

24 servers turned down our handshake because they only speak the newest version of the protocol, 2026-07-28. They're running, so we count them as working. But not every client can reach them yet. The current release of the official TypeScript SDK (1.32.2) doesn't speak that version, and the Python SDK (2.3.0), which speaks both, listed the tools of 16 of them in our test. The others asked for a header it doesn't send, or answered with a tool list it rejects. If you run a server like this, accepting the older handshake too keeps today's apps working.

Test yours

From a terminal, send the first message every client sends. Put your server's address in place of the example:

curl -i https://your-server.example/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"check","version":"1.0"}}}'

A working server answers 200 with JSON that includes "protocolVersion", either as the body or on the data: line of an event stream. Anything else is what an app sees too. To walk through the whole exchange, run the official MCP Inspector: npx @modelcontextprotocol/inspector.

Or test it in your browser with our free MCP server test, the same check this report used. For a human look, send us the address: we tell you what fails and why, free, within 24 hours. If you have an error message instead, our MCP connector looks it up in our fix guides from inside Claude or any MCP client.

How we measured

The list. Every server in the official MCP Registry on 19 November 2026 that lists a remote address: 45,531 entries. We kept the ones published under a company's own domain, not a GitHub account or a free hosting subdomain, that answer without the user's own API key. Then one address per host, leaving out platforms that host thousands of servers for others, such as Smithery, Apify and Glama. That leaves 21,180 company servers.

The check. The first two steps any client takes: initialize, then list the tools, over Streamable HTTP, with 10 seconds for each. For servers on the older SSE transport, opening the event stream. We never call a tool.

Twice. A server counts as broken only if both checks failed, at least 10 minutes apart. When the address didn't resolve, we also asked Cloudflare's and Google's public DNS, and counted it only if both failed too.

What counts as working. Any real answer from the server, including one that asks for a key or a payment (401, 402, 403), limits our rate (429), blocks our region (451) or speaks only the newest protocol version.

Left out of the count. 102 addresses on throwaway developer tunnels (trycloudflare.com, ngrok and similar) and 99 servers switched off on purpose: 410 Gone, or a host saying the deployment is disabled or paused. They failed too, but nobody expects them to answer.

Limits. We check from one place, in Canada, so a server that blocks our network, or was down for exactly those minutes, counts as broken. And a server that passes can still fail when a tool is called, which we don't test.

No names. We don't publish which servers failed. The counts behind this page are in 2026-11.json. We wrote the checking scripts and this page with AI tools, including Claude. Every number comes from the check logs.

Next edition: December 2026.

Most of these breaks take minutes to fix, once someone knows they're there.