Docs
How AgentBlock works
AgentBlock turns one decision — which known AI agents may fetch your pages — into a deployable blocking kit. This page documents the method: the database, the evidence standard, the policy model, and what the kit can and cannot do.
The known-agent database
Every kit is generated from a versioned, dated database of known AI agents (current version: 2026-10-18). Each entry names the agent, its vendor, its user-agent token, its category (for example training, retrieval, or user-action), and an evidence label with the source and last-verified date, all shown on the database page.
Evidence labels are strict:
- verified — the vendor publishes this user-agent token in its own documentation.
- reported — seen in reputable bot directories but not published by the vendor; labeled as such, never upgraded silently.
- watch — agents with no published user-agent. Watch entries are never guessed into blocking rules.
The policy model
On the builder you set one policy per agent: allow, block, or leave unlisted. Agents you have never heard of classify as UNKNOWN — a first-class answer, not a silent allow. Your policy is the only input; the kit is generated deterministically from it and the database version, so the same policy on the same database version always produces the same kit.
What you deploy
The kit ships in three forms generated from the same policy: a Cloudflare Worker, a WordPress plugin, and a plain JavaScript snippet for other stacks. Supporting tools check your work after deployment: the audit and verify-kit tools confirm which agents your live site actually admits, and the stale check flags kits built on an older database version.
Honest scope
AgentBlock blocks known agent user-agents. User-agent blocking can be bypassed by spoofing, and agents that publish no user-agent cannot be blocked this way — they stay watch entries. AgentBlock is a policy enforcement layer for declared agents, not a general bot firewall.