Sentinel is the pre-action risk Skill an onchain agent calls before it signs. It reads the target cold (bytecode, storage, ownership, token state) and returns a safe / caution / dangerous verdict with reasons.
Name an address. Sentinel judges it.
Sentinel is the pre-action risk Skill an onchain agent calls before it signs. It reads the target cold (bytecode, storage, ownership, token state) and returns a safe / caution / dangerous verdict with reasons.
The gate below is that Skill running live. Drop in an address (try the Backdoor) and you see exactly what an agent sees when it calls Sentinel.
Every verdict is built from read-only Foundry cast reads of the live chain — no indexer, no third-party API.
Sentinel pulls the target's bytecode, proxy storage slots, and token calls — all read-only, all through Foundry.
It walks the opcodes and proxy slots for the patterns that actually drain wallets — not surface heuristics. Each one found subtracts from the safety score.
Every check starts at a perfect 100; each signal subtracts from it, so multiple small risks pull the safety score down.
A band turns the score into a clear decision — safe, caution, or dangerous — an agent can act on directly.
Never just a number. Every verdict carries the human-readable why — the audit trail behind the call.
The verdict becomes an action — approve or block, plus a de-risked size — so value moves only when it's safe to.
Six contracts deployed on Atlantic, each tripping a distinct signal. Sentinel reads them live, straight from the chain.
Read-only, keyless, reproducible — the most-called primitive in any on-chain agent stack.
Run a check ▸