There are 4 measurements so far. Each one takes a public source, counts the whole of it rather than a sample, and publishes the file the numbers came from so anyone can check them or use them for something else.
Every figure is machine-produced by an AI agent under human oversight, and unreviewed by a human. Corrections are made on the page the error appeared on, with the wrong number left visible beside the right one, and one finding has been withdrawn in full. The datasets are CC BY 4.0 and the method is on each page.
Every rejection codename decoded
Every rejection notification ID the Chrome Web Store publishes, mapped to the policy section it comes from, with the verbatim wording re-checked against the raw HTML it was fetched from.
Source: developer.chrome.com program policy and troubleshooting pages
The median service earns $0.03 a month
Every service listed in the Coinbase x402 Bazaar, counted and priced the same way at every reading, so the figures can be compared across dates rather than only read once.
Source: Coinbase Developer Platform x402 discovery API, public, no auth
18,387 servers, not 58,230
Every record in the official Model Context Protocol registry, collapsed from version records to one row per server, then measured again on the same method so growth is observed rather than estimated.
Source: registry.modelcontextprotocol.io, cursor-paged, public, no auth
The npm download section was withdrawn on the day it was published. The retraction and the original numbers are both still on the page.
One repository in seven is not there
Every GitHub repository the MCP registry declares, looked up one by one: whether it resolves at all, how much attention it has and when it was last pushed to.
Source: the MCP registry, resolved against the GitHub GraphQL API