vadgr-computer-use

Local MCP server for desktop automation. 13 tools for capture, mouse, keyboard, and platform introspection. The calling agent takes a screenshot, reasons over the pixels, and drives mouse/keyboard through the server.

Tested with Claude Code, Codex CLI, and Gemini CLI (same server, same tools, same prompt).

Platforms: works on Linux (X11 and Wayland incl. GNOME 46-50, KDE, wlroots), Windows native, WSL2, and macOS. On Linux run vadgr-cua install-deps once after install (clipboard backend + input permissions); on macOS grant Accessibility + Screen Recording on first run. See First run on Linux, First run on macOS, and Platform support.

Install

pip install vadgr-computer-use

That ships a console script called vadgr-cua. On Linux, run the one-time system-dependency step (the second of the two install commands):

vadgr-cua install-deps        # prints the plan; add --yes to run it

It provisions the clipboard backend (wl-clipboard) and /dev/uinput access via a single graphical auth prompt (pkexec, falling back to sudo). pip can't install those — they are OS packages — so this command bridges the gap. See First run on Linux.

Verify:

vadgr-cua doctor
# Linux also reports the resolved capture/input backends under "platform_backends"

On WSL2, the bridge daemon auto-launches the first time a tool is called. On other platforms it's a no-op; direct backends handle everything.

Wire it into your agent

Pick your client. The server command is vadgr-cua --transport stdio in every case. Each agent launches that stdio process itself, so it needs the full path to the binary unless vadgr-cua is already on the agent's PATH.

First, find the path:

which vadgr-cua
# global install: /home/you/.local/bin/vadgr-cua
# venv install:  /path/to/.venv/bin/vadgr-cua

Substitute that path in each config below.

Claude Code

Project-level (.mcp.json at the repo root you want to automate from):

{
  "mcpServers": {
    "vadgr-computer-use": {
      "type": "stdio",
      "command": "/path/to/vadgr-cua",
      "args": ["--transport", "stdio"]
    }
  }
}

User-level (add to ~/.claude.json under mcpServers with the same shape).

Verify: claude mcp list should print vadgr-computer-use: ... ✓ Connected.

Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.vadgr-computer-use]
command = "/path/to/vadgr-cua"
args = ["--transport", "stdio"]

Verify: codex mcp list should list vadgr-computer-use with status enabled.

Gemini CLI

gemini mcp add --scope user --trust \
  vadgr-computer-use /path/to/vadgr-cua \
  -- --transport stdio

That writes ~/.gemini/settings.json. Verify by running an interactive session: Gemini shows MCP tool calls inline.

Try it

Once the wire-up is done, any of these commands launch the client, which starts vadgr-cua --transport stdio in the background via MCP, and drives your desktop. Same prompt, same tools: pick the client you already use.

Sanity check (focus + Ctrl+A):

Take a screenshot, tell me in one sentence what application is in focus,
then press Ctrl+A and take another screenshot to confirm the action.

Claude Code

Interactive (most common):

claude --dangerously-skip-permissions
# then paste the prompt at the > cursor

Headless one-shot:

claude --dangerously-skip-permissions -p \
  "Take a screenshot, tell me what app is in focus, then press Ctrl+A and screenshot again."

Codex CLI

Headless one-shot (the usual way to drive Codex):

codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check \
  "Take a screenshot, tell me what app is in focus, then press Ctrl+A and screenshot again."

Expected output (abbreviated):

mcp: vadgr-computer-use/screenshot (completed)
mcp: vadgr-computer-use/key_press (completed)
mcp: vadgr-computer-use/screenshot (completed)
The focused app is <...>; Ctrl+A selected its content.

Gemini CLI

Works end-to-end, but pixel grounding on full-screen shots is weaker than Claude/Codex: first-attempt clicks on small targets can miss by 20-60 px (the model usually recovers via screenshot_region crops). Pass the model explicitly, since the default may silently fall back to an older Gemini on some accounts:

gemini -m gemini-3.1-pro-preview -p \
  "Use only vadgr-computer-use tools. Take a screenshot, tell me what app is in focus, then press Ctrl+A and screenshot again." \
  -y --allowed-mcp-server-names vadgr-computer-use

Fuller example: play a song on YouTube Music (Codex)

A Chrome window is already open with a "YouTube Music" tab. One call:

codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check \
  "Use only vadgr-computer-use MCP tools. In the already-open Chrome,
   switch to the YouTube Music tab, search 'Space Oddity David Bowie',
   and play the first result."

Real transcript (trimmed):

mcp: vadgr-computer-use/screenshot (completed)
mcp: vadgr-computer-use/click (completed)        # YouTube Music tab
mcp: vadgr-computer-use/click (completed)        # search box
mcp: vadgr-computer-use/type_text (completed)
mcp: vadgr-computer-use/key_press (completed)    # enter
mcp: vadgr-computer-use/click (completed)        # first result
mcp: vadgr-computer-use/click (completed)        # dismiss ad overlay
mcp: vadgr-computer-use/screenshot (completed)   # verify now-playing bar
Yes, "Space Oddity" by David Bowie is now playing.

How it works

The LLM owns the "where to click" decision; the server owns "how to click it precisely". No other abstraction in between.

Platform support

The Linux backend is selected per session by a capability resolver (run vadgr-cua doctor to see what it picks and why):

pip install vadgr-computer-use pulls jeepney and python-xlib automatically on Linux (pure-Python, no compilation). The pixel-input fallback uses a pure-Python /dev/uinput writer, so no C compiler is needed; the optional evdev-backed path is available via pip install vadgr-computer-use[linux-uinput]. The clipboard backend (wl-clipboard) and /dev/uinput access are OS-level and installed by vadgr-cua install-deps. Foreground-window detection on Wayland uses AT-SPI2 if available; install with pip install vadgr-computer-use[linux-atspi] to enable it.

On macOS, pip install vadgr-computer-use pulls pyobjc-framework-Quartz and pyobjc-framework-ApplicationServices (wheel install, no compilation). No Homebrew packages required.

First run on Linux

After pip install, run the one-time system-dependency step:

vadgr-cua install-deps --yes   # one pkexec/sudo prompt; omit --yes to preview the plan

It installs the clipboard backend (wl-clipboard) and sets up /dev/uinput access (udev rule + input group) under a single graphical auth prompt. pip cannot install these because they are OS packages, not Python wheels.

On GNOME 49/50 Wayland, the first screenshot() shows a one-time GNOME consent dialog (the XDG Screenshot portal); click Share and the grant is remembered, so later screenshots are silent. For unattended/remote runs, trigger one screenshot while you are at the machine first so the prompt is out of the way. On GNOME 46-48, gnome-screenshot is used and there is no prompt. Input (mouse/keyboard) on GNOME uses Mutter RemoteDesktop and needs no prompt.

Check what the resolver selected:

vadgr-cua doctor
# "platform_backends": { "capture": {"selected": "portal"}, "input": {"selected": "mutter-remotedesktop"}, ... }

First run on macOS

You can pre-grant permissions before connecting an agent:

vadgr-cua setup

That fires the Accessibility and Screen Recording prompts and prints the current grant state as JSON. Toggle the entries on in System Settings when prompted. If you skip this, the same prompts fire on the first MCP tool call from your agent.

The first time the MCP server captures the screen or injects an input event, macOS opens System Settings to two panes and asks you to grant the running Python interpreter:

• Privacy & Security -> Screen Recording (required for screenshot() / screenshot_region()).
• Privacy & Security -> Accessibility (required for clicks, typing, scroll, drag).

Toggle both for the python binary that runs vadgr-cua (e.g. /path/to/.venv/bin/python or /opt/homebrew/bin/python3.12). The grant is per-interpreter and persists; you will not be asked again. Verify status:

vadgr-cua doctor
# {... "macos_accessibility_granted": true, "macos_screen_recording_granted": true,
#      "python_executable": "/opt/homebrew/bin/python3.12" }

Apple enforces these prompts at the OS level for every screen-capture / input-injection API; they cannot be skipped.

If you later revoke either permission in System Settings, the next MCP tool call detects it via CGPreflightScreenCaptureAccess() / AXIsProcessTrusted(), opens System Settings to the right pane, and returns a structured error to the agent. Toggle the entry back on and the next call works. No silent black screenshots, no hunting through System Settings.

If the WSL2 daemon can't start (e.g. no Windows Python available), the server falls back to a slower PowerShell path. See Daemon management below.

MCP tools (13)

Capture (2)

• screenshot(): full screen, downscaled to CU_MAX_WIDTH (auto-picks 1024 / 1280 / 1366).
• screenshot_region(x, y, w, h): cropped region.

Input (8)

• click(x, y) / double_click(x, y) / right_click(x, y)
• move_mouse(x, y) / drag(start_x, start_y, end_x, end_y, duration=0.5)
• scroll(x, y, amount): positive = up, negative = down
• type_text(text) / key_press(keys): keys like ctrl+s, alt+tab, enter

Platform info (3)

• get_platform() / get_platform_info() / get_screen_size()

Daemon management (WSL2)

Most users never touch this. For when you do:

vadgr-cua doctor           # JSON: platform, Windows Python, daemon state, port, hash
vadgr-cua install-daemon   # Eager deploy + launch
vadgr-cua stop-daemon      # Kill the running daemon
vadgr-cua restart-daemon   # Stop then start

The daemon file is deployed to %USERPROFILE%\vadgr\daemon.py and listens on TCP 127.0.0.1:19542. After pip install -U vadgr-computer-use, the next MCP session detects the version-hash drift via a ping handshake and redeploys the daemon automatically.

Library usage

from computer_use import ComputerUseEngine

engine = ComputerUseEngine()
shot = engine.screenshot()
engine.click(500, 300)
engine.type_text("hello")

The library is just the input/capture primitives, no LLM or agent loop inside. To drive it with a model, point an MCP client (Claude Code, Codex, Gemini, or your own) at the vadgr-cua server as shown above.

Environment

Tests

pip install -e ".[dev]"
pytest computer_use/tests -q

License

Apache 2.0. See LICENSE.

Part of Vadgr

• vadgr: workflow engine (brain)
• vadgr-computer-use: desktop automation MCP (eyes)
• vadgr-agent-os: containerized agent runtime