mcpcensus

Check whether an AI agent could discover and connect to an MCP server for a domain — the same probe that produces MCP Census.

npx mcpcensus check example.com

It asks only for things a domain owner published on purpose: a short, versioned list of .well-known paths, a DNS TXT record, conventional endpoints, text fallbacks, and — only if those already found an endpoint — one unauthenticated JSON-RPC call. It never invokes a tool, never sends credentials, never tests for a weakness, and honours robots.txt for every path.

What the result means

A negative is not proof of absence. The probe covers a published candidate list; a server somewhere else is a server it missed, and it says so. A refusal — 401, 403, 429, 5xx — is recorded as inconclusive, never as absence, because it is a fact about the probe rather than about the domain.

Missing OAuth protected-resource metadata is never a finding: authorization is optional in MCP, so a public unauthenticated server correctly publishes none.

The rest

• Methodology, including the limitations — several of which describe the project's own mistakes
• Crawler ethics, and how to opt out
• The open dataset, CC-BY-4.0

Code is Apache-2.0. Radixia is measured in its own dataset like everyone else.