Your drafts, your voice profiles, and your people live in a database on your Mac. The server runs the model and forgets. Here is the whole data path, plainly.
An assistant that learns how you write to each person only works if you can trust where that learning lives. So here is Cue’s data path, end to end, in plain words.
What stays on your Mac
- Your voice profiles, one per conversation, learned from replies you keep
- Your drafts and their history
- Your people, memory facts, and tracked promises
- Your settings, commands, and personal dictionary
All of it sits in a local database you can open and read. Deleting the app deletes the learning. There is no account state to export from us because we never had it.
What crosses the wire
To draft, answer, or transcribe, Cue sends the minimum the model needs: the text or audio of the moment, over TLS, to our proxy, which holds the model API key and meters usage. The model runs. The response streams back. Nothing is written down on the way through. Screenshots are on by default so answers use what is on screen, can be turned off anytime, and follow the same rule.
There is no history of your drafts, screens, or messages to leak, because we never keep one.
The honest trade-off: metering means we count requests, so rate limits and fair use work. Counts are numbers, not content. That is the entire list of what our servers know about you.
Local-first is not a settings toggle we added late. It is why the app is fast, why it works the same on a plane with cached profiles, and why the privacy page is short.