Manual Only
Analysis and advice. It reads, explains and recommends. It drafts nothing and does nothing.
How it works
The setup is deliberately short. Earning your confidence is deliberately not — the product is built so that widening its authority is something you do gradually, on evidence, and can walk back without penalty.
One sign-in, one permanent handle. Your first workspace is created for you — there is no blank canvas to stare at.
Pick a role template or write your own, give it a mission in one line, set its permissions, and connect a mailbox. Minutes, not an afternoon.
New mail is triaged against its mission and your history. Routine threads get drafted; anything needing judgement is surfaced with a one-line reason.
Approve or edit each draft from one queue. As it matches your style, you widen what it handles — on your schedule, reversible at any point.
Analysis and advice. It reads, explains and recommends. It drafts nothing and does nothing.
It drafts and stages. You approve every send. This is the ceiling — autonomous execution is not part of this release.
A colleague you cannot rely on when things break is not a colleague. These are the failure paths, and what each one does instead of losing your mail.
The daily report is generated deterministically from the event log, so it still arrives. Drafting pauses and says so, rather than producing something worse.
Everything you have is still readable — mail, memory, search. Sends composed offline queue automatically and dispatch on reconnect, reporting the result back.
It surfaces on a dedicated attention screen with the plain-language cause and a one-click fix, instead of failing silently in the background.
You edit before approving, and the correction becomes part of what it knows. Nothing was sent, because nothing is ever sent without approval.
Notifications are pooled into digests, each worker has a daily cap, and quiet hours are respected. High-risk items break through all three, because that is what they are for.