The inbox is loaded and the agent is idle. Run it, and every move it wants to make lands here for your final approval before anything executes.
Operation mode
Decide how much the agent can do on its own — and where it must stop and ask you.
Confidence threshold
Auto-approval only fires above this; anything less waits for you.
85%
No mode ever auto-sends, auto-forwards, or clears the risk line. Autonomy moves the routine — never the consequential.
Held actionsnothing proposed yet
Nothing is waiting on you.The agent hasn't touched the inbox — it can't, until you run it, and not without your sign-off on every move.
Audit trail
Every proposal and decision, logged.
Waiting for the first run.
Mock inbox, real model. The summary and the full text under each item are the email's own words; the purple block is the agent's read of it — kept separate on purpose, so you can check one against the other. The agent triages on the summary, the way a fast first pass would; you review with the whole thing. It proposes, never executes. Opened locally without a backend, it runs on a baked fallback set.