CrawlCheck
Registry · MCP · monthly

MCP reliability report, 2026-10

Month to date, rebuilt about once a day; the edition is sealed after the month ends. What the remote MCP servers listed in the official registry did when CrawlCheck contacted them: how often they answered the handshake, how fast, and how often their tool lists changed. No tool is ever called and no server is named here.

Partial month: observations start 2026-10-08, when the panel began.

Endpoints observed6,000 of 6,000 in the panel, from 4,985 publishers
Observations24,624, of which 24,508 answered (99.5%, 95% interval 99.4–99.6%)
Asked for credentials10,376 of the answered observations
Went down11 endpoints, 11 times (two failed observations in a row after answering)
Time to answermedian endpoint 768 ms; all answers p50 1,075 ms, p95 4,625 ms
Tool lists changed396 of 3,495 endpoints (11.3%), 508 changes; mean 12.7 tools

Availability per endpoint

Endpoints with at least 5 observations this month: 1,902. Another 4,098 had fewer, and 0 never answered this month.

100%1,87398.5%
99 to 100%00%
95 to 99%00%
80 to 95%221.2%
50 to 80%40.2%
Under 50%30.2%

How failed observations failed

unreachable77
http error35
redirected4

Method and limits

each panel member is contacted in rotation from CrawlCheck's network (Cloudflare): initialize, notifications/initialized, tools/list, DELETE session; no tool is ever called. Answered = the handshake completed or the endpoint asked for credentials. Availability per endpoint is answered / observations inside the month.

  • one vantage network: an endpoint unreachable only from Cloudflare reads as down
  • the panel holds endpoints that answered at least once, so endpoints that never answered are outside it
  • endpoints are contacted in rotation, so short outages between two observations are not seen
  • no server or publisher is named in this report; per-server readings are in the reliability index

Per-server readings: MCP reliability index. Signed JSON: /api/v1/mcp/reliability/report?month=2026-10.