Raw data

Every figure on the State of the MCP Registry page comes from these files. They are published so the arithmetic can be re-walked by someone who does not trust it, which is the only version of “carefully measured” that is worth anything.

File	What it is
census-2026-07-30.json 1.27 MB	Full-registry sweep, 2026-07-30. All 10,716 active-latest remote URLs, each with a health class and HTTP status. This is the baseline every later comparison is anchored to.
successor-probe-20260803.json 830 KB	The successor probe, 2026-08-03. For every 7/30 URL that was no longer the active-latest entry on 8/2, both the old endpoint and its current successor probed in the same run. Contains the resolved pairs, the full 2×2, the four contract-continuity rows, and the raw probe result for all 1,217 URLs.

Method

Endpoints are probed anonymously over HTTP with an MCP initialize handshake (protocolVersion 2025-06-18) followed by tools/list. A server counts as answering only if tools/list returns a result. Servers requiring credentials return 401 and are counted as not answering — the right reading for “can a caller use this”, the wrong one for “does this server exist”. Both readings are recoverable from the per-URL error codes in the files.

Known limits

Anonymous probes only. Auth-gated servers are indistinguishable from broken ones.
Single observation window per run. A 503 may be a deploy, not a death.
Successor resolution uses the registry’s name field, which is self-declared by the operator whose endpoint moved. Contract continuity — the stronger witness — could only be checked for 4 pairs, and is reported as individual rows rather than a rate.
Counting unit matters: some servers register many versions at trivially different URLs (one gateway appears 76 times, differing only in a query string). Figures are deduplicated to distinct server name unless stated otherwise.