Let agents act freely within safe limits, and require a human to mint authority for the risky stuff.
Model the approval boundary as authority. Give the agent a standing warrant scoped to low-risk actions (read, small spend). For anything above the line — a large payment, a destructive operation — the agent has no warrant, so it must request one. A human (or a policy service) reviews and issues a one-time, tightly scoped warrant for that specific action.
A boolean "approved" flag is easy to spoof or forget to check. A warrant is a signed capability: the action simply can't happen without one, the approval is recorded in the audit log, and the one-time warrant expires immediately after use.
Liceat is permission infrastructure for AI agents. You issue a warrant — a scoped, spend-capped, revocable token — that grants an agent narrow authority; the agent presents it at the point of action; Liceat enforces it (scope, spend, rate, holder-binding, revocation) and records every decision in a tamper-evident log. It is open source with a zero-dependency SDK for TypeScript and Python.
The SDK and self-hosted gateway are free and open source. The hosted gateway at liceat.com gives 10,000 free decisions with no card, then meters at fractions of a cent per decision — payable by card or USDC.
Free to self-host. 10,000 free decisions on the hosted gateway — no card.
Get an API key Read the docs