What leaves your Mac

Written for whoever has to approve the install — a security lead, a lawyer, or you reading the network policy before a client call.

Your speech never leaves the machine

Recognition, translation and style rewriting all run on your Mac. No audio and no transcript is ever uploaded, on any plan. Nothing is kept after a session either: the app stores your transcripts locally and there is no server behind them to store anything.

That is not a promise about intent. It is a consequence of the design: there is no service to send audio to.

Every address the app can contact

Nothing here carries your speech or your text.

Your account. Signing in, signing up from the first-run setup (email, password, and a signature made with the install key), and asking for a token. These go to our own server.

Usage counters — only with consent. If you allow statistics, the app sends numbers, never text. What is in them: how many dictations and in which language or language pair; how long you spoke; how many characters came out — the count, never the characters; how long recognition, translation and styling each took; which paid feature was blocked; and your country, taken from the system region. Two doors: by session when you are signed in, and when you are not, by a stable device id plus a signature made with the install key.

If you did not allow statistics, none of this is sent at all.

Update check — only when you press the button. The app reads verbaide.com/version.json and compares one number. It never checks on its own.

Translation dictionaries from cdn.verbaide.com. The catalogue is fetched when Settings are open, or when the language screen of the first-run setup is open. The files themselves are fetched when you download them. Both requests carry no cookies and no authorization header. Only the ru/uk→en dictionary ships inside the app; every other direction, including en→ru/uk, is downloaded.

Speech models from huggingface.co. Two of them: the main Parakeet TDT model, pulled the first time it is needed — usually on the try-it screen of the first-run setup — and a smaller CTC model (98 MB) used for the term dictionaries, installed in the background on first dictation.

Style-model metadata from cdn.verbaide.com/intelligence. Catalogue when Settings are open, manifest when a download starts. Metadata only — the weights are not on that host.

Style-model weights from huggingface.co/Qwen/…. Downloaded by the Mac of whoever chose to install the local model, the same way the speech model is.

What the free tier does

One download of the speech model — without it even free dictation does not work. After that: no account and no limits.

Two requests do continue, and they do not depend on your plan: the translation-dictionary catalogue and the style-model catalogue, both fetched while Settings are open. Neither carries cookies, an authorization header, or anything about you. Close Settings and they stop.

Where to check this yourself

Point a proxy at the app and watch. Every address above is in CLAUDE.md in the source tree, and the network policy that decides whether a request may leave at all lives in the Rust core, in one list shared by macOS and Windows.