How Parrot labels your mail, and exactly what leaves your account

Every message that arrives gets exactly one category: work, personal, finance, travel, social, updates, or important. This is how that happens, stated plainly, because the version of this post we published first was not accurate and we would rather correct it in public than quietly delete it.

What actually happens

When a message arrives, we send three things to a model: the sender, the subject, and the first 1500 characters of the body. The model returns one word — the category — and that is the entire exchange. No tools, no history, no other message.

It is a small, cheap call on purpose. Sorting mail does not need a reasoning model, and it does not need your whole thread.

What we keep

The label, and the message id it belongs to. Not the text. The body we sent is not stored by us, and under our agreement with the provider it is not retained on their side, not logged for human review, and not used to train anything.

Your mailbox itself is never copied. It stays on Gmail, Outlook, or your IMAP server, and Parrot reads it live over their API.

The claim we withdrew

An earlier version of this post said categorization ran on-device, on a quantized classifier shipped with the client, and that corrections adjusted local weights. That described a design we prototyped and did not ship. The shipped product makes a network call, and several pages on this site repeated the local-model claim for longer than they should have. They have been corrected.

If you chose Parrot because sorting was local, that is a real reason to reconsider, and we would rather you know now.

Where the line actually falls

The honest boundary is not local versus remote — everything AI in Parrot is remote. It is how much leaves and how long it lives. Labeling sends a sender line and a truncated body. Drafting sends the thread being answered. Learning your writing voice sends short excerpts of your own sent mail, and most of that system does not call a model at all: ten of the thirteen things it measures about your writing are counted arithmetically in plain code.

Every one of those calls is stateless. Nothing is kept, and nothing you write becomes training data.