FAQ

Does anything leave my computer?

No. Speech recognition, the language and vision models and the screenshots all run and stay locally. There is no account and no telemetry. The only network use is downloading a model when you ask for one, and the one deliberate exception is pointing a model endpoint at a server on another machine, which the settings app flags in amber.

Do I need a GPU?

No. Utter works with no models at all for rule-matched commands, and speech recognition runs on CPU with whisper.cpp. A GPU makes speech faster and is needed for the optional decision-head and vision models. The comfortable setup for all three is a 16 GB or larger NVIDIA card; see Models for the recommendation per GPU class.

Why did Utter go to sleep by itself?

By default Utter sleeps after 15 minutes without being used, so the graphics card is free while you work on something else. Only Utter activity keeps it awake: a push-to-talk key, a spoken command, or waking up. Typing or clicking in other apps does not count, and it never dozes off while it is listening or carrying out a command. Hold either push-to-talk key to wake it; speech is back in a fraction of a second and the bigger models reload in the background. Turn it off or change the time under Sleep when idle on the General page, or set [sleep] on_idle and idle_minutes in the config file.

Does it work on X11, GNOME, KDE, Hyprland, sway?

Utter targets Wayland. niri and KDE Plasma (KWin) are first-class: a small compositor layer detects which one is running and uses its native interface for window and workspace actions (niri msg on niri, KWin’s D-Bus interfaces on Plasma). Other Wayland compositors get partial support — the generic tools it uses (wtype, ydotool, grim, wl-clipboard, PipeWire, MPRIS, AT-SPI) work across compositors, so dictation, typing and launching work, but compositor-specific actions (focus, move, workspaces) are not implemented for them. X11 is not supported.

Windows or macOS?

macOS is experimental — native Speech.framework + whisper.cpp voice, Quartz push-to-talk and injection, and an unsigned .dmg; it has not been tested on real Apple hardware. See macOS. Windows is not supported and has not been started.

What is the difference between the assistant key and the dictation key?

Hold the assistant key and your words become an action. Hold the dictation key and your words are typed into the focused field, nothing more. Each has its own sound and waveform colour. See Getting started.

Why Insert and F13 as defaults?

Because the reference setup remaps Right Alt to Insert and Caps Lock to F13 with keyd, which gives two large, comfortable keys that nothing else uses. Change them on the Voice page to any key you like.

Can the AI run arbitrary commands?

No. Rules go first. When no rule matches, a small local model chooses among fully resolved actions Utter has already prepared; it cannot write its own. Running terminal commands at all is off by default, and anything derived from screen content can never become an argument. See Trust and safety.

Will curl | bash install things without asking?

Not on its own. Piped input is not a terminal, so the wizard cannot prompt: a bare pipe prints the plan and exits without changing anything. Add --yes to accept the recommended defaults. You can also read the script first; it is install.sh at the root of the repository.

Does the installer download models?

Never on its own. The models step asks per tier and only pulls when you opt in and provide a source. The settings app also downloads nothing until you press the button.

How do I uninstall?

./install.sh --uninstall from a checkout, or the remote one-liner with --uninstall. Your config and models are kept unless you remove them yourself. See Install from the web.

Can I use my own language-model server?

Yes. Anything OpenAI-compatible works: vLLM, Ollama, llama.cpp. Set [router] llm_base_url and llm_model, and [vision] base_url and model, or use the LLM and Perception pages.

How do I teach it a new app?

Create a YAML profile in ~/.config/utter/profiles/ or edit the app on the App actions page, then regenerate the catalog. See Apps and actions.

Which languages does it speak?

The settings app and the installer ship 10 UI locales: English, Spanish, German, French, Italian, Portuguese, Chinese, Japanese, Korean and Russian. The spoken language is a separate setting (it defaults to your system locale), so you can use a Spanish UI and speak English. Speech recognition depends on the model: the default is an English-only (.en) whisper model, and you opt into a multilingual model (small, large-v3-turbo, …) from the settings app or the installer. Utter never downloads a model on its own.

Is Utter finished?

No. It is young software built with AI coding assistants and reviewed by a human. Read the code before trusting it with anything important, and report anything that looks wrong.