Every directory can tell you how many servers it lists. Listing is free; answering is not. These numbers come from fetching each endpoint on a schedule and keeping the changes, so they describe what responds rather than what was submitted.

108,054
resources indexed
from 57,214 publishers
86.9%
of probed endpoints answer
61,361 answering, 9,246 not
35.3%
of introspected servers expose tools
3,714 require credentials
99,376
verified tools
median 7 per server that answers
354 ms
median response
of endpoints that answer
93,202
recorded changes
across 66,392 endpoints in 19.24 days

Churn

6,008 endpoints stopped answering and 7,932 started again, within the 19.24-day window above. counts transitions inside the window above, not a rate. Multiply nothing by 365: a short window over a young index says what happened, not what happens.

How this is measured

Every endpoint in the index is fetched on a schedule. `answering` means it responded at all, not that it responded correctly, and never that it is safe or good. Tool counts are read from each server's own tools/list, so a server behind credentials reports as auth-required rather than as empty. Only state CHANGES are stored, so a stable endpoint costs one row rather than one per probe.

What these numbers are not

Take the data

This page as JSON: https://neuronto.com/state-of-mcp with Accept: application/json. The underlying liveness observations, including the list of endpoints that stopped answering, are at https://neuronto.com/liveness and https://neuronto.com/liveness?dead=1. Free to use, redistribute and build on, no attribution required, no key. If you run a registry, the dead list is the useful half.

Changes as they happen: /feed.