The official MCP registry by the numbers: 37,628 servers read, November 2026
We connected to every remote server the official MCP registry lists and read what each one hands an agent. How many answer, how many want a key, what they tell the model, and what our tool poisoning rules found.
On 20 November 2026 the official MCP registry listed 37,628 remote servers. We connected to every
one of them the way any MCP client does, with initialize and then tools/list. We read what
came back. We never called a tool. These are the numbers from that read, and they come from our
live registry report, which updates as we read.
The registry grew by a quarter in under three weeks
On 1 November the registry listed 29,629 remote servers. By 20 November it listed 37,628. That is 7,999 more in 19 days, or 27%. Each new server is a new thing an agent can be told to connect to.
Who answers
- 23,600 servers (63%) listed their tools to anyone who asked.
- 10,422 (28%) wanted credentials first. We count these, but we cannot read them.
- 3,389 (9%) were down, gone, or not an MCP server at that address.
What they hand an agent
The servers that answered listed 376,689 tools. The median server has 7. Each tool comes with a name, a description and a schema. An agent reads all of it to decide what to call.
17,407 servers also set instructions. That is about three in four of the servers that
answered. Instructions are free text a client puts in front of the model when the server connects. The user has
not asked for anything yet. Most of it is plain help: call this tool first, ask before you spend money. It is
also the channel a poisoned server would use, so it is worth reading.
What our rules flagged
11 servers were flagged. These are the rules that fired, and one server can trip more than one. Three had a schema field named for the chat or a secret, with no word on why the tool needs it. Two had a block in their tool text dressed up as a system prompt. Two had a parameter that asks the agent to send secrets or the chat. One told the agent to read local secrets or its own config. One tried to change the agent's role. One told the agent to hide what it does from the user. A server is also flagged, with no rule, when something it names is on a public list of known bad hosts.
A person reads every flag. We did that on 4, 9 and 19 November, and we have not yet found a confirmed attack. Most flags were security tools whose own descriptions name the attacks they look for. On 19 November, 12 flags came from one tool that a single publisher copied onto 12 servers. A few we cannot settle from what the server shows us, so they stay flagged until we read them again. When a rule is wrong we change it, and we say so on our corrections page.
What these numbers cannot tell you
- A server that wants credentials is counted, not checked.
- Local servers that run over stdio are not in the registry's remote list, so they are not here.
- A server can change its tools at any time. Each figure is as of its read. We read each server again about every two weeks. See how often servers change their tools.
- A clean read means these rules found nothing. It does not mean a server is harmless.
Look up a server, or check one yourself
Every server in this read has its own page in our MCP server directory, with its tools, its setup and the date of our last read. Each page has a feed that tells you when its tools change. To check a server that is not in the registry, paste its URL into the free MCP server check. Our first, smaller read of the registry is what 2,760 MCP servers tell your agent.
Check your lockfile for known malware. Free, no card.
Lazaretto is the pre-install checkpoint agents call themselves. Try it · pricing · connect your agent.