MCP Census: can an agent find your MCP server?

Per-domain measurements of whether an AI agent could discover and connect to an MCP server for each organisation in a frozen, reproducible population, starting from nothing but the domain name. The population is 7,422 organisations the official MCP Registry proves run a server. That choice is deliberate and stated: these have already done the hard part, so the figures here are an upper bound on the wider web rather than a sample of laggards. Each deposit contains the per-domain rows with full evidence, the frozen input universe with its provenance and download date, a flattened CSV, a DuckDB script for Parquet, and a summary whose every aggregate declares its own denominator. Earlier dated releases are kept unchanged alongside, because a citable artifact that changes is not one. What the data does not say. A negative is not proof of absence: the probe covers a published, versioned candidate list, and a server elsewhere is a server it missed. A refusal (401, 402, 403, 429, 5xx) is recorded as inconclusive rather than as absence, because it is a fact about the crawl. Missing OAuth protected-resource metadata is never non-compliance, since authorization is optional in MCP. Conformance is for the specification's authors to determine; this measures, it does not judge. Conflict of interest. The authors' own domain, radixia.ai, is included in the measured population rather than excluded, and this is disclosed at the top of the methodology. Code is Apache-2.0 and lives in the same repository. The data is CC-BY-4.0, which is the licence of this record.