computer-use-linux MCP
computer-use-linux is an MCP server with an on-screen agent activity indicator that lets user clicks reach the application.
- 816 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent can view the desktop, click controls, type text and change application data.
Install
- Works in
Manual install
npm install -g @agent-sh/computer-use-linuxInstalls the npm CLI, which also starts the MCP server with the `mcp` subcommand. This is the README’s first option for a CLI and any stdio client.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent can view the desktop, click controls, type text and change application data.
This is third-party code. Review the repository files before installing.
What it does
computer-use-linux is an MCP server connecting an agent to a Linux graphical session. The server reads an accessibility tree: interface elements with roles and names. The agent uses it to select a window, click controls and type text. After acting, the agent retrieves application state again to see the changes. The server supports X11 and Wayland, with input and window targeting depending on the desktop. The project also includes a CLI and native Pi tools. An on-screen activity indicator lets user clicks pass through to the application.
Who it is for. Linux desktop application testers assigning an agent work with forms, windows and interface controls.
Good fit when
- A tester checks a desktop form through accessible controls.
- Before acting, the agent captures the selected window.
- After changing Wayland sessions, a user checks input readiness.
- The agent compares application state before and after a click.
Not a fit when
- A browser-specific tool is sufficient without full desktop access.
- The agent must not affect the user’s mouse, keyboard or focused window.
- Every input method must work identically on every compositor.
Example request
Check Linux desktop readiness and locate the test application window without clicking or typing yet.MCP
- Transport
- stdio
- Authentication
- not required
Limitations
Prebuilt binaries require glibc 2.28 or later; older systems need a source build. Input requires components for the current graphical session, such as xdotool on X11 or a desktop portal or uinput access on Wayland. Plugin installation does not add the CLI to the system path, so terminal commands require a separate CLI installation. The server operates the user’s real desktop: it does not create a separate, isolated environment for agent actions.
How to disable. Check the result in the target window before disconnecting. Finish any drag and release held keys and mouse buttons so subsequent input does not change application data. Then disable the server in the MCP client or remove the plugin.
FAQ
How do I install the Claude Code plugin?
Run `claude plugin marketplace add agent-sh/computer-use-linux`, then `claude plugin install computer-use-linux@computer-use-linux`.
What does the plugin download on first start?
The plugin downloads pinned release binaries and checks SHA-256. To use your own build, set `COMPUTER_USE_LINUX_BIN` in the client environment to the executable’s absolute path; relative paths are ignored.
Is there a Rust installation route?
Yes. `cargo install computer-use-linux` builds and installs the published crate.
What happens when window selection is ambiguous?
The server rejects the call. Refresh the window list and choose a unique target. Retrieve the identifier again after a window title changes.
Can Pi connect without MCP?
Yes. `pi install npm:@agent-sh/computer-use-linux` adds native lazy-loaded Pi tools.
Does window targeting work in stock labwc?
No. It requires a specific labwc IPC patch; the stock build lacks the socket the server uses to target windows.
Discussion
No questions or reviews yet.
Related
MCP serversOpen Computer Use
open-computer-use
An open MCP server for desktop control: the agent reads app state and clicks, types and scrolls on macOS, Linux and Windows
An MCP server for driving Windows: the agent opens apps, clicks and types, reads window and UI state, and works with files and PowerShell
Clawd Cursor: desktop control through MCP
Clawd Cursor
An MCP server that lets an agent drive windows and apps: it reads the screen as structure and acts by element name
MCP server and wincli CLI: agents drive Windows apps by control name through UI Automation instead of screen coordinates
A Swift MCP server that lets an agent drive any macOS app: click, type and read the interface tree
An MCP server for macOS: the agent moves the mouse, types, manages windows and captures the screen when an app has no API
README in short
GNOME, KDE and other desktops have their own documented window-targeting methods. The author maintains an isolated desktop as a separate project outside this server.