agent-cage

agent-cage is a standalone, fail-closed safety library for AI agents, written in Python and released under the MIT license with no dependencies outside the standard library. A single guard() call returns allow, ask, or deny on every proposed action. Policies are frozen at load time so running code cannot silently widen permissions. Irreversible steps require human ticket approvals, an irreversible kill switch halts the agent when tripped, budget caps are enforced in a thread-safe manner, and every decision is recorded in a hash-chained, HMAC-signed audit log. The library ships with 133 clean-room tests, all passing, and a self-commissioned adversarial audit found 15 issues (4 rated critical); all were remediated and re-verified. The project is early-stage with growing adoption; this application seeks retroactive recognition of shipped work and a maintenance runway, not funding contingent on future features.