Box — a native GTK4/libadwaita app for running AI locally (chat, voice, vision, document Q&A)
I'm the developer. Box is a native Linux desktop app for running AI models fully on-device — not Electron, not a browser shell. It starts in under a second and integrates with the GTK desktop. No account, no cloud, no telemetry.
Features:
- Local chat with Gemma 4 models via Google's LiteRT-LM runtime — streaming, Markdown + LaTeX
- Voice — voice messages + hands-free voice conversation mode (offline Piper TTS, six voices)
- Live camera vision — webcam frames sent to a vision model
- Document Q&A (RAG) — index PDFs/docs/images per-chat or into reusable Notebooks; answers cite their sources
- Web search + workspace-scoped filesystem tools, plus a multi-step agent mode
- Persistent memory across chats
- Six themes — Catppuccin and Dracula — with accent colours and bubble styles
- Every capability is a separate opt-in switch, off by default, with permission prompts before any tool acts
Install — a .deb for Ubuntu (amd64) is on the Releases page: sudo apt install ./box_0.1.0_amd64.deb. System dependencies are pulled automatically; models download on first run, then it's fully offline. CPU-only is fine.
Honest notes:
- Built for Ubuntu 26.04 (amd64); depends on Python 3.14 + GTK4/libadwaita.
- Closed-source — the
.debships compiled code. (My Android app is open source; this one isn't.)
GitHub: https://github.com/jegly/Box