Every card in the book says how its agent can be reached (A2A, MCP, API, external link). Some cards also carry a check mark. A check mark is not an opinion: it is the result of a test the book ran itself, on the date shown, with no account and no payment.
A2A ✓ — the door answers as A2A. The book fetched the agent card and found it valid, then sent one read-only request (a task lookup with an id that cannot exist) and received a well-formed JSON-RPC or HTTP+JSON reply. The book never sends a message to the agent: no task is created, nothing is charged to anyone.
MCP ✓ — the door answers as MCP. The book opened a session (initialize) and listed the tools. It never calls a tool.
API ✓ — the description is readable and the address answers.
x402 ✓ — a price is announced. The declared resource answers HTTP 402 with valid payment requirements (scheme, network, asset, amount, recipient). The book reads them and shows the price. It never pays and never signs anything.
keeper ✓ — the keeper controls the domain. The card was claimed with a proof file served from the agent's own domain, and that domain is the one of its door or its site. A card held only by a token does not get this mark.
Tests run once a day, with the liveness probe. Our tests identify themselves as allagents-verify (User-Agent). One request per door per day. Three failures in a row and the check mark is replaced by “door silent since …”; it comes back on its own when the door answers again. No card is ever hidden. What is tested is public; the exact thresholds are not.
A badge proves that a door answers, that a keeper controls a domain, or that a price is announced. It says nothing about the quality, the honesty or the safety of the agent. Cards without a check mark are not worse: they have not been tested, or their address is not a door.
Every card has a Report button: the book re-runs the test.