About

agentprobe.org is an independent, read-only census of the agentic web.

What it does

Every public registry, directory and well-known convention that lists AI agents, MCP servers or agent catalogs is harvested in full. Every remote endpoint those listings point at is probed with the smallest real protocol interaction that proves it is what it claims to be — an MCP server/discover or initialize, an A2A agent card and one message, an ARD catalog fetch, an OpenAPI document, an unauthenticated /v1/models. Outcomes, latencies, protocol eras and capabilities are recorded every cycle; the time series is the product.

Why

Registries publish what operators submit. None of them can credibly certify that the entries they profit from listing are alive, and none measure each other. The result is a discovery layer where a six-figure number of "agents" is quoted and a small fraction answers. A neutral instrument that says which — with denominators, dates and an open method — is the missing piece, in the same way independent certificate authorities and SSL Labs were for TLS.

What it is not

Not a registry: nothing can be submitted here. Not a directory to browse: the human pages exist to verify and compare, the machine surface exists to be called. Not a ranking of quality: "conformant" means the handshake worked and the capabilities were as declared, nothing more.

Use the numbers

Aggregates on this site are CC BY 4.0: quote them with "agentprobe.org, date" and the denominator. Per-endpoint pages are permalinks. Bulk data is not public; contact the operator for research access.

Be excluded

Block the User-Agent agentprobe/ in your robots.txt; it is re-read every run. Endpoints are probed at most once per cycle, one request in flight per host, with an identifying User-Agent and contact URL.

Operator

Venture Inventory GmbH, Berlin. Source, probe specification and design brief: github.com/parismic/agentprobe.