Finally, an MCP server that admits Linux has more than X11

What it does

computer-use-linux is an MCP server and CLI that turns a local Linux desktop into an instrument an LLM agent can play. It reads AT-SPI accessibility trees, takes bounded screenshots, and drives clicks, scrolls, and keystrokes across GNOME, KDE, Hyprland, i3, and COSMIC. Any MCP host—Claude Desktop, Codex, Hermes Agent, or your own client—can spawn it to observe or mutate live application state.

The interesting bit

Instead of assuming every Linux session is just X11 with extra steps, the project chains compositor-specific APIs in order of preference—GNOME Shell extensions, KWin DBus, Hyprland sockets, i3 IPC, COSMIC helpers—so window targeting works out of the box on both Wayland and X11. Input falls back through org.freedesktop.portal.RemoteDesktop, wtype, and ydotool until something sticks, which means the agent can usually type Unicode without knowing which display server is running.

Key highlights

• Wayland-first input via desktop portals, wtype, and ydotool, with X11 handled as a best-effort fallback.
• Compositor-aware window targeting that probes GNOME, KWin, Hyprland, i3, COSMIC, and generic X11/EWMH, then reports exactly which backend won.
• Semantic AT-SPI selectors (role, name, text, states) let agents click and type without brittle pixel coordinates.
• A built-in doctor tool returns a single JSON readiness report so hosts can surface platform blockers without parsing logs.
• MCP ToolAnnotations classify every tool by mutability and destructiveness, and the optional run_shell tool is gated behind an explicit environment flag.

Caveats

• Manual validation is currently limited to Ubuntu 25.10 with GNOME; other compositor backends are implemented and tested at the contract level, but real behavior depends on the session exposing its expected API.
• The optional run_shell tool is explicitly not a sandbox; enabling it grants the caller the same file and network authority as the user running the server.

Verdict

Worth a look if you are building an LLM agent that needs to drive a real Linux desktop—especially on Wayland—without writing a new backend for every compositor. Skip it if you are macOS-only or need a fully sandboxed, zero-trust automation layer.