6,026 sites implement Model Context Protocol
of 8,817 tested · measured 2026-08-23
8,817 hosts strict-re-graded, 23 Aug 2026, of 14,351 listed
Checks passed across the 6,026 probed hosts: Markdown negotiation 5,114 · Markdown twin at a predictable URL 4,500 · llms.txt 5,826 · API catalog 918 · MCP server card answers 6,026.
What these sites let an agent do
The tools most often registered, by how many of the 6,026 register each. Real verbs, read from the page at runtime — enumerated, never called.
997 distinct tool names across 6,026 sites.
How much of this is one plugin
Two instruments read this set and they do not agree, which is the finding rather than a problem. The tool signature — the exact set of names a site registers, treated as a fingerprint — puts 17% of the 6,026 in a cluster with at least one other site, and the largest single cluster is 5.1%: one WordPress plugin, registering the same two tools on every site that installs it. The platform column — which fingerprints hosts — attributes just 89.1%, because it does not fingerprint plugins at all. Where the two disagree the signature is the better instrument, and the size of that gap is why we built it.
6,026 pass when we test the behaviour. 2,791 do not.
Any directory can list sites that claim a standard — the claim is in the source and costs nothing. Separating the two requires testing the behaviour, and dating the result. That is the only thing on this page you cannot get elsewhere.
What Model Context Protocol is
MCP is the protocol an agent uses to discover and call a server's tools. On the open web a site advertises one through a server card, and an agent connects to the endpoint the card names. Read the MCP specification — it is the authority on the design and we are not trying to be.
How we test it
We fetch the server card and then CONNECT to the endpoint it advertises, and count it only where the server answered tools/list — or challenged us for auth, which is an answer.
A server card is a file. It costs nothing to publish and proves nothing about whether anything is listening. The gap between cards published and servers that answer is the single largest gap we measure.
Sites that implement it
84% of these are on mintlify, 5.2% of these are on teachify. The platform supplies the machinery; the domain signals the commitment, and a report says which applies.
All 6,026 of these sites, through the API: every one, in a fixed order, 25 per call, with the total stated.
Model Context Protocol is one signal among several. The rubric says exactly what we look for and how each tier is reached, and a probe can be requested — the probe decides, never the request.
Method and date
Measured 2026-08-23. 8,817 hosts strict-re-graded, 23 Aug 2026, of 14,351 listed Every figure on this page carries the population it was measured against; a count without a denominator is not a measurement. How the index is built · All standards.