Statistics
What MCP servers return when you call them
Every figure on these pages came from a connection we made to a real address on a real date. Nothing is estimated, modelled, or inferred from download counts, and every page states the population it measured and the day it looked.
The population
The MCP registry census
We called 1,100 of the remote addresses the official registry lists, once each. 13% did not answer as an MCP server.
Edition 2026-08 · swept 20 Aug 2026
The sample
The endpoints we hand-picked
64 public endpoints we chose and checked ourselves, in depth: tool names, field frequencies, and which revision each one negotiated.
66 checks · 24 July 2026 to 4 October 2026
Why there are two of them
They answer different questions, so their numbers will not match and are not meant to. The census sweeps the whole registry once, shallowly, which is the only way to say anything about how much of a catalog is reachable. The hand-picked corpus is a much smaller set we check more closely, which is the only way to say what a tool list actually contains. Any figure quoted from either page should say which one it came from.
Check one yourself
The MCP connection tester runs the same check against a single address and gives you a dated record you can share. Built a server of your own? Launch it on the weekly board.