How it works

Connect Microsoft 365 or Gmail by OAuth. On Microsoft the grant is Mail.ReadWrite and Mail.Send; on Google it is one scope, gmail.modify. Both are the narrowest permission the loop can actually run on: write access is what lets an approved reply go out threaded inside the customer’s own conversation rather than as a fresh message wearing “Re:”.

Nothing moves. There is no forwarding address for customers to learn, no second place for support to live, and if you disconnected tomorrow your inbox would be exactly where you left it.

The pass keeps a watermark — the newest message it has recorded — and asks your provider only for unread mail that arrived after it, up to 60messages in one run. That is a deliberate design rather than an optimisation: Microsoft’s own throttling guidance names continuous polling as a pattern that gets applications throttled, and a mailbox that gets throttled is a mailbox that stops sending.

The watermark advances only once a message has actually been written down, so a pass that dies halfway re-reads rather than steps over what it never got to. Re-reading is free — the same message recorded twice is dropped on a unique key.

Every message gets three things:

• A category — question, bug, billing, sales, onboarding, feature request, complaint, legal, cancellation, automated or spam.
• A priority — urgent means production is down, money is moving wrongly, a deadline is today, or the sender is escalating.
• Evidence — the verbatim lines from the message that drove the label. A quote the message does not literally contain is dropped before it is ever shown, because an invented quote on an evidence line is worse than no evidence at all.

Bounces, out-of-office replies and cold pitches are recorded and then done. They never become a card — a morning queue full of postmaster notices is not a queue.

The draft may only draw on the thread itself and the answers your team has already saved. Where the answer is in neither, it says plainly that it is checking and will come back rather than filling the gap. Three things stop a bad draft before you ever see it: one carrying a bracketed placeholder is thrown away, one that cannot be parsed is thrown away, and one that commits to a refund, a credit, a discount, a guarantee or a date the thread never mentioned is thrown away and the thread goes to a human instead.

Complaints, legal threats and cancellations are never drafted at all. That is a branch in the code rather than a confidence threshold, so it cannot be tuned away by a model having a good day.

What lands in the morning is one card per conversation: the drafted reply, what the sender is asking for, and the evidence. Approving does not send it — it stages the reply for 120 seconds, and the dispatcher refuses to pick a row up before that moment has passed. Undo inside the window flips it to cancelled and the customer never saw anything.

Two caps bound the whole loop, and neither can be waived: a thread past 3 agent replies goes to a human and stays there, and past 120replies in a day the queue simply stops draining until tomorrow. Nothing is lost — the direction of failure is always “too quiet”.

Each read, classification, draft, handoff, approval, edit, kill, send and cancellation is written to a ledger with the evidence it acted on. The receipt is written before the mail rail is touched, so a send that could not be logged does not go out at all — a receipt you cannot produce is the same as an action nobody authorised.

Want to see it on a real morning? See the live demo →