Guardrails
A per-session safety net: your agent asks before risky actions — you decide.
- e.g. a force-push or reading a secret
- The agent asks before acting
- From your phone, wherever you are
- Approved actions proceed; denied ones stop
Guardrails give every session a safety net for risky agent actions. Each category — reading secrets, destructive commands, protected branches, and outward or irreversible actions like force-push, publish, or deploy — can be set to Deny, Confirm, or Off. Guardrails are on by default (Confirm): the agent asks first, and you approve or deny from your phone. They’re built to catch accidental risky actions — not a hard security boundary.
Every category starts on Confirm, so the agent asks before anything risky — no setup needed.
Tune each category for each session: Deny blocks the action outright, Confirm asks you first, Off trusts the agent.
Confirmation requests reach your phone — approve or deny without being at your desk.
Designed to catch accidental destructive actions. It’s not a hard security boundary — use environment-level restrictions when you need hard guarantees.
- 1Open the session’s ⋯ menu → Guardrails
Guardrails are configured per session, from the session menu.
- 2Set each category
Choose Deny, Confirm, or Off for secrets, destructive commands, protected branches, and outward actions.
- 3Save
The settings apply to the session right away — the agent asks (or is blocked) from the next risky action on.
FAQ
Related
Ready to try it?