The MCP Census

The MCP Census is a public-evidence checking service for Model Context Protocol servers. A lookup gathers observable facts about a server's identity, repository, package, hosted endpoint, maintenance, and registry presence, then reports a result such as PASS, REVIEW, BLOCK, or UNKNOWN. It is meant to be a first gate before an agent installs, enables, or calls a server. That addresses a practical problem in the fast-growing MCP ecosystem. A directory entry may help someone discover a tool, but it doesn't by itself show whether the listed package still exists, the remote endpoint is alive, or a familiar name really belongs to the vendor it appears to represent.

The service is careful about the boundary of its verdicts. It describes itself as a public-evidence gate, not a malware scanner, penetration test, compliance certificate, or guarantee that a server is safe in production. PASS means the selected published policy passed against the facts and source times in the result. REVIEW means a required fact is missing, stale, inconclusive, or needs human judgment. BLOCK marks a confirmed hard failure under the policy. UNKNOWN means the Census has no conclusive row for the subject, not that the server has been cleared. Users still need to inspect requested permissions, source code, scopes, data access, and runtime isolation.

A person can search and inspect servers on the website, while developers and agents can use REST or the hosted MCP endpoint. Preflight applies one named, versioned policy to one exact canonical server and returns ordered reasons with evidence details. The available built-in policies vary in how strongly they require first-party identity and hosted liveness. A compact stamp exposes the same kind of policy decision for automation. The API also supports search, lookup, credit balance checks, policy discovery, recent records, coverage statistics, and direct server resolution. An OpenAPI description and human reference are published for teams building their own integration.

One of the stronger workflow features is whole-config auditing. A client can submit an MCP configuration and receive one verdict line for every entry, including an exit code suitable for scripts. The audit recognizes common configuration shapes used by Claude Desktop, Cursor, VS Code, Codex, and standard mcpServers files, resolving entries by remote URL, package, or canonical name. Keyless use includes one configuration of up to 25 entries per IP each day. With an API key, an audit costs one credit per configuration per day and supports up to 100 entries. A strict option can make REVIEW or UNKNOWN produce a distinct nonzero result, while any BLOCK can fail a continuous integration check.

The Census also treats server status as something that changes after installation. Its unmetered change feed reports observed transitions such as a registry appearance, a hosted endpoint going down or returning, a changed tool list, an endpoint move, or a verification and health update. Clients poll with a cursor so they only receive changes since their previous request. Paid or credited users can subscribe a server to a signed webhook watch. The project publishes a GitHub Actions workflow and a short agent rule intended to run the check before each connection. Together, the audit, change feed, and watches make it possible to keep the gate in a recurring workflow instead of treating lookup as a one-time review.

Pricing is usage based but has meaningful free access. Anonymous visitors can make ten unique searches or lookups per day per IP, while several statistics and change endpoints are unmetered. A free account includes 1,000 credits per UTC month with no card. Search and lookup use one credit, config audit uses one credit per configuration per day, preflight or stamp uses five, and subscribing a watch uses ten. Repeating the same billed subject on the same UTC day is free. The Builder plan costs $19 per month for 2,000 monthly credits, a $9 pack provides 1,000 credits that don't expire, and the Team plan costs $49 per month for 10,000 credits.

The MCP Census fits developers, platform teams, and agent operators who want a consistent evidence check in front of MCP connections without building their own registry monitor. It stands apart from discovery catalogs because the product is the identity and health decision rather than hosting, ranking, or sign-in. The reasons and public facts make each result inspectable, and both REST and MCP access make it practical for automation. Its limitation is equally important. Public evidence can catch dead repositories, deprecated packages, liveness changes, and identity gaps, but it can't inspect every behavior that will occur after connection. Used as the first decision point in a layered review, it can reduce avoidable mistakes without pretending to replace code review or sandboxing.