I got tired of all the desktop agent tools being macOS only, so I built one for Linux.

Like many of you, I’ve been messing around with OpenClaw (formerly clawdbot) and the whole "vibe coding" concept. It's cool, but finding a decent tool that actually drives the UI on Linux was a pain. Everything seems to be Mac-first right now.

Since I do all my local inference on Linux, I built a dedicated tool for it.

It's called Peepbo : https://github.com/LichAmnesia/peepbo

Basically, it's a lightweight Node/TS wrapper that connects your local VLM (LLaVA, Qwen-VL, etc) to your desktop Linux environment.

How it works:

• Vision: Wraps scrot, gnome-screenshot, or gdbus so the model can see the screen.
• Control: Uses xdotool to handle mouse/keyboard inputs.
• Wayland: Yes, it works on GNOME Wayland, but you'll need to run in unsafe mode (details in the readme).

It's open source. Give it a shot if you're trying to build agents on Linux and let me know if it breaks anything.