A dictation tool hears everything you say. That is an enormous amount of trust, and the most honest way to earn it is to make the scary version impossible by design.
Your audio is transcribed and discarded. We do not keep recordings, we do not build voice profiles, and there is no archive of what you said sitting on a disk somewhere. The transcript — text — is what matters, and even that lives on your device, not on our servers.
Keys and account credentials live in the operating system’s secure store on your computer, never in our database and never in logs. The desktop app talks to the backend with a license token, not your secrets.
Features that read your screen for context are off by default, and when on they send only a short, bounded snippet — never the whole screen, never anything we retain. You decide, per session, whether that trade is worth it.
The boring architecture is the point. There is no interesting breach to have when there is nothing kept to take.