root@tr0p:~#Home

Atticus

Press a hotkey, speak, get a spoken answer. On your own Linux desktop.

Atticus puts a small always-on-top panel in the corner of your screen. It transcribes Brazilian Portuguese or English with faster-whisper, answers in a Piper neural voice, and gates every command the model runs behind a bubblewrap sandbox.

  • platformOmarchy · Hyprland
  • licenseMIT
  • stackGTK4 · faster-whisper · Piper
  • speechlocal

InstallSource on GitHub

tr0p@omarchy: ~/apps/atticus-linux
# bound to Ctrl+Alt+Space: press, speak, press again
$ atticus-ctl toggle
heard    "coloca um timer de 10 minutos para o macarrão"
instant  timer · 10:00 · macarrão (no AI round trip)

# anything else goes to the model, behind the gate
heard    "what is on my screen?"
sandbox  bwrap · files read-only · no network
[+] answer streamed to the panel and read aloud
$ 

Features

A desktop voice assistant that stays out of the way, keeps your speech on your machine and never lets the model run a command unchecked.

Overlay that never steals focus
A corner panel built on wlr-layer-shell with GTK4 and WebKitGTK. It floats above your windows and leaves the keyboard where it was.
Hotkey or wake word
Push to talk with a keybinding, or turn on the always-on wake word. Conversation mode reopens the microphone when Atticus ends on a question.
Local speech, both ways
Recognition tuned for Brazilian Portuguese with faster-whisper, English supported. The Piper voice stays loaded so answers start speaking at once.
Instant commands
Timers, alarms, reminders, volume, brightness, media, lists, Wi-Fi, Bluetooth and lock screen run locally in Portuguese and English, with no AI round trip.
Sandboxed by default
Model commands run inside bubblewrap: files read-only, keys and browser profiles hidden, no network. Real changes need an Allow card. rm, sudo and git push never run.
Sees and drives the desktop
Screen reading with grim and tesseract, web search, local file search, native click and typing, plus smart home through a Home Assistant server.

Usage

Bind one key and talk. Atticus runs as a systemd user service, so you manage it like any other.

Talk to it hotkey · wake word

Press Ctrl+Alt+Space, speak, press again to send. The answer streams into the panel and is read aloud. Close with Esc or let it fade.

$ atticus-ctl toggle
$ hyprctl reload

Run it as a service systemd user

Settings live in ~/.config/atticus/settings.json. Restart the service after editing them and follow the logs live.

$ systemctl --user status atticus
$ systemctl --user restart atticus
$ journalctl --user -u atticus -f

Voice pipeline

Every request takes the same path from your voice to a spoken answer.

  1. Trigger A hotkey or the wake word opens the microphone.
  2. Transcribe faster-whisper turns speech into text, locally.
  3. Route Everyday commands run at once. The rest goes to the model.
  4. Gate Model commands pass the sandbox or wait for your Allow.
  5. Answer The reply streams into the overlay panel.
  6. Speak Piper reads it aloud in a neural voice.

Numbers

MetricValue
Speech languagesPT-BR / EN
Command safety tiers4
Default Whisper modelsmall
Telemetrynone

Install

Clone and build. build.sh sets up the virtual environment, downloads the Portuguese voice and starts the user service.

$ git clone https://github.com/marcostolosa/atticus-linux.git && cd atticus-linux && ./build.sh
  • Requires Omarchy or another Hyprland / Wayland session, and Python 3.11+.
  • build.sh checks the system packages and prints the exact pacman line for anything missing.
  • The brain defaults to the claude CLI. An OpenAI or Claude-style endpoint (Ollama, LM Studio) works too.
  • Log out and back in once after the first install so ydotool can click and type.