On 2026-06-10, of 1631 MCP registry servers probed, 592 (36.3%) answered an anonymous MCP initialize handshake.
571 (35.0%) were alive but auth-gated (reachable, need credentials — not dead). 325 (19.9%) returned a 404 at the exact URL the registry advertises, and 92 (5.6%) were unreachable. The registry lists 11557 servers in total; 4915 expose a remote endpoint, of which we have probed 1631 so far (33% coverage).
| Live (handshake OK) | 592 (36.3%) |
| Alive but auth-gated | 571 (35.0%) |
| Speaks MCP (init error) | 8 |
| Reachable, non-MCP | 12 |
| Endpoint not found (404 at listed URL) | 325 (19.9%) |
| Unreachable | 92 (5.6%) |
| Protocol error (other) | 31 |
Tamper-evidence: this run's hash 204cbf7e05a01f76… chains to the prior run GENESIS… (SHA-256). Machine-readable: /atlas/census.json.
Each run reads the population of the official Model Context Protocol registry
(registry.modelcontextprotocol.io/v0/servers) and, for every server that lists a
remote endpoint, sends one anonymous MCP initialize JSON-RPC handshake from
Cloudflare's network with an 8-second timeout and the User-Agent
Dominion-Census/1.0 (+https://dominionobservatory.com).
Responses are categorized without conflating distinct outcomes:
Limits and honesty. This measures reachability of the registry-listed endpoint at probe time, not server quality, safety, or correctness. Servers behind allow-lists may show as unreachable from our network. Stdio-only servers (no remote endpoint) are out of scope. The registry is in preview and its contents change. This is an independent, unofficial project and is not affiliated with the MCP project, Anthropic, GitHub, or Microsoft.
Opt out. A server operator who does not want to be probed or listed can email the contact on our site and we will exclude the endpoint.
Numbers are hash-chained run-to-run (SHA-256) so the time-series is provably contemporaneous and not backdated. The machine-readable dataset is at /atlas/census.json.