AI Assistant For Local Use

Lounge
1

Hi all,

I’ve been building Box for Ubuntu, an AI assistant that runs on your own machine. There’s no account to create, and it sends nothing over the network unless you switch on a feature that needs to. It’s a native GTK4 / libadwaita app on top of Google’s LiteRT-LM runtime, so it opens in about a second and stays light on memory. You download a model once, then it works offline.

It handles the usual chat, but it can also search the web, read and audit your files and logs, answer questions from documents you give it, and look through your webcam. Turn on agent mode and it strings those together to get through multi-step jobs.

What it does

  • Local chat with on-device models in .litertlm format. Gemma 4 E2B and E4B are the ones I’d start with; both go up to 128K context. Output streams as the model generates it, Markdown and LaTeX render in place, and you can run on CPU or GPU.

  • Web & file tools + agent mode. Searches DuckDuckGo over HTTPS, no API key. Reads and writes files inside a workspace folder you pick. Agent mode chains tool calls for multi-step jobs, and each call shows up as a card with its arguments and result so you can see what it did.

    • File & log audit. Hand it a log or config and ask it to check for problems, e.g. “audit /var/log/dmesg for anything suspicious.” It reads the whole file, past the model’s context window if it has to, and hands back one report.

    • Knowledge base. Attach PDFs, Markdown, source, or text and ask questions against them; the reply shows which passages it pulled from. Notebooks are reusable document sets you index once and attach to whatever chats you want.

    • Voice. Offline TTS through Piper, six voices. A hands-free mode lets you talk, hear the reply out loud, then it listens again. The model takes the audio directly instead of working off a transcript.

    • Live camera vision. Capture a frame for the model to look at, or leave Vision Mode on so it grabs one each turn. Runs through GStreamer/PipeWire with a V4L2 fallback, respects the camera portal, and turns the light off when you finish.

    • Persistent memory. It remembers what you tell it to remember, in a store you can open and edit.

    • Themes. Catppuccin (Mocha, Latte, Frappé, Macchiato), Dracula, and Dracula Pro, with accent colors, bubble styles, and custom fonts.

    Privacy model Every feature is its own switch and they all ship off. Vision, audio, TTS, knowledge base, web search, filesystem, agent mode, memory: you turn on what you want. Anything that touches your machine asks first, with allow-once, this-chat, always, or deny. File writes and deletes ask every time. File access stays inside a folder you choose. Network requests have to use HTTPS.

    check it out here github.com/jegly/B0x

1 Like
2

Moved to the Lounge which is a good place for general information and discussions.

I deleted your other post since it was essentially a duplicate of this one.

Thanks for sharing with the community.

2 Likes