tusk
open source · Apache-2.0

The open voice shell for your desktop and your AI agents.

Talk, and the desktop does it. Always listening, no wake word, any model.

Get Tusk on GitHub Watch the 90-second demo
Linux + GNOME/X11 today. Needs Docker and a Groq API key.
listening
you
“open gedit and type a poem”
> gedit opened, poem typed
you
“move it left”
> window moved
room
“…so, lunch at one?”
ignored · not for Tusk
you
“start pair-coding”
> function written in your editor
you
“use gpt-oss-120b”
> model swapped by voice

See it work

90 seconds, hands off the keyboard.

A cheap model listens, a capable model acts.

mic > VAD > STT > gatekeeper > agent > MCP adapters
01 · listen
A cheap gatekeeper model listens continuously and filters ambient speech, deciding whether you were talking to Tusk at all.
02 · act
A capable agent acts through MCP tools on your real desktop: apps, windows, typing, dictation, pair-coding in your own editor.
03 · swap
Swap models by voice, mid-session, and plug in other agents as the backend.

Understands plain speech. Acts on the real desktop.

acts on the real desktop text or chat only
LLM understanding tusk Wispr Flow, Superwhisper, OS assistants
Fixed grammar Talon, AutoHotkey Classic dictation
Open source, model-agnostic, works with any editor.

Who it's for

Developers with RSI or motor impairments who need to work without typing all day.
Keyboard-optional developers pairing with AI coding agents who'd rather keep their hands off the keyboard.

Have RSI and work on Linux?

I'm looking for five design partners: Linux developers who use voice daily and want to shape what ships next.
Become a design partner

Install

Linux + GNOME/X11 today.
$ git clone https://github.com/vovka/tusk.git
$ cd tusk
$ make demo
Requires Docker and a Groq API key. Full setup, configuration and troubleshooting are in the README.
Get Tusk on GitHub

Status and roadmap

Q4 2026
Public repo, safety confirmation layer, measured latency
Q1 2027
Hosted voice stack beta, Wayland, memory across sessions
Q2 2027
Windows adapter
H2 2027
macOS, team policy and audit, adapter marketplace
Working prototype, bootstrapped, Apache-2.0. Known gaps: Linux/X11 only, no destructive-action confirmation yet.
Built by Volodymyr Shcherbyna in Ukraine, a daily Talon voice-control user.