Where it runs. The service runs on our own server in the UK, the same one that hosts our customers' websites. It connects to your mailbox over an encrypted IMAP connection with a password or app password you give us, stored encrypted; the key that unlocks it lives on the server, not in the database.
What the model sees. The text of each email is sent to Anthropic's Claude API over an encrypted connection to be sorted and, where you have asked for it, drafted. Under Anthropic's commercial terms that text is not used to train their models. Their servers are in the United States, which we state plainly because some customers need to know it.
What we keep. A log of each email's sender, subject, bucket, summary and draft, with the tokens and cost, so you can see exactly what it did and what it cost. The email itself stays in your mailbox and is never copied anywhere else by us.
The one-pager is the only source of facts. The drafter writes from your one-pager (what you do, your hours, what you never promise by email) and your own past replies. It cannot look anything else up, which is exactly why it leaves brackets rather than guessing.
The demo at /ai/try/ keeps the verdict and the draft so we can improve it, and does not keep the email you pasted.