Your voice, turned into actions on your desktop.

curl -fsSL https://utter.sujaisubbanna.com/install.sh | bash

In a terminal this opens an interactive wizard. Piped without —yes it only prints the plan and changes nothing. Linux on Wayland, x86_64.

Push-to-talk

Hold the assistant key and speak, or hold the dictation key to type what you say into any field.

Desktop actions

Open apps and websites, focus and close windows, switch workspaces, control media, press app shortcuts.

Rules first

Most commands match fast, predictable rules. A small local model only chooses among prepared actions. It never invents commands.

Offline

Speech recognition, AI models and screenshots all stay on your machine. No account, no telemetry.

Frees your GPU

Say “go to sleep”, or just stop using it for a while, and the models unload. Hold a key to wake it in a fraction of a second.

Two modes, two keys

Hold the assistant key (Insert by default) to run a command. Hold the dictation key (F13 by default) to type what you say. Each mode has its own sound and waveform colour.

Understands context

Utter knows which app is focused and what is on screen. It uses accessibility information first and takes a screenshot only as a last resort.

Per-app actions you can edit

More than 100 app profiles with their shortcuts. Change any key combination from the settings app or a YAML file.

Safe by default

Terminal commands and raw input stay off until you turn them on. Important actions ask first. Screen text can select, but never author, an action.

IntroductionWhat Utter is and how it decides what to do.

ConfigurationHotkeys, speech, AI and vision models, audio, profiles.

Trust and safetyProvenance, confirmations, dangerous actions and permissions.

Writing a pluginSpeak the JSON-RPC protocol from Python or Rust.

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. Apache-2.0.