Privacy and your data
Kudu is built so that what you say stays on your computer.
What Kudu never does
- Send audio or text to a server. Transcription and formatting run on your computer.
- Ask for an account, or collect analytics, telemetry or crash reports.
- Keep audio. It's held in memory while it's transcribed, then discarded.
- Record your keystrokes. On Linux Kudu reads the keyboard only to notice its own shortcut.
When Kudu uses the internet
Only to download a model you chose: the speech model on first launch and, on Windows and Linux, the optional smart-formatting model. Models come from GitHub (speech) and Hugging Face (language model). After that, Kudu works offline.
What's stored, and where
| What | Why |
|---|---|
History (history.jsonl) | Your dictations, for the Home page and stats. Kept 90 days by default; clear it any time. |
Usage totals (usage.json) | Words and time per day for Insights. No text. |
| Dictionary, snippets, settings | Your words, snippets and preferences. |
Log (kudu.log, Windows and Linux) | Errors and timings, to diagnose problems. Never dictated text. |
| System | Folder |
|---|---|
| macOS | ~/Library/Application Support/Kudu |
| Windows | %APPDATA%\Kudu (settings, history) and %LOCALAPPDATA%\Kudu (models, log) |
| Linux | ~/.config/kudu and ~/.local/share/kudu |
Reading the screen
When you start dictating, Kudu reads the visible text of the window you're in, through your system's accessibility interface, for a fraction of a second, and keeps only names and terms, in memory, to spell them right. After a dictation it may read the text field for up to a minute to learn from your corrections; only the corrected word is saved. Both can be turned off in Settings.
The full policy is in the repository: PRIVACY.md.