What it does
Jev answers six narrow risk questions, while deterministic policy keeps authority over the decision. The bundled simulator can use a fake provider; live judgments need a TypeSafe key.
Agent tooling
ActionGate reviews an agent's proposed tool call before allowing it, blocking it or asking for human approval.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Jev answers six narrow risk questions, while deterministic policy keeps authority over the decision. The bundled simulator can use a fake provider; live judgments need a TypeSafe key.
Start by writing down risky tasks your automated helper might try, such as sending money refunds. List clear rules and spending limits for each task. A developer can set up this program to check those limits before any action happens, pausing questionable requests until a person says yes.
You can try this guard system first with pretend tools that cannot touch real funds. To run real safety checks, your developer connects an access key from TypeSafe or OpenRouter. Because outside security experts have not reviewed this early code yet, try it only with practice accounts.