Computer Use (CLI)

Use this skill to drive and validate a Linux GUI from the command line:

• Take screenshots to observe the current UI state
• Click / type / scroll / drag to exercise flows
• Smoke-test desktop apps (Electron / Tauri / React Native Desktop) inside a sandboxed X11 session

It’s most useful when you need GUI automation but only have shell access (no native UI automation framework available).

Quick start

1. Ensure dependencies are installed (see “Dependencies”).
2. Point the commands at the active display (usually already set):
   • export DISPLAY=:1 (or whatever your container uses)
3. Take a screenshot to establish the agent’s visual state:
   • cu-screenshot --base64
4. Perform actions in API coordinate space (the resized screenshot), then screenshot again:
   • cu-click --x 512 --y 384 --base64
   • cu-type --text "hello" --base64

Tip: start the app under the same display, for example:

(DISPLAY=:1 ./my-electron-app &) && sleep 2
cu-screenshot --base64

Coordinate spaces (critical)

These commands support two coordinate spaces:

• API space (default): coordinates refer to the resized screenshot (e.g. XGA 1024×768). This keeps vision tokens lower and matches “computer use” style interactions.
• Screen space: coordinates refer to the real display resolution.

By default, if scaling is enabled, inputs are treated as API space and are scaled up before calling xdotool.

Use:

• --space api (default) or --space screen
• cu-info to see screen size, chosen API size, and scale factors.

Standard interaction loop (recommended)

Use this loop to keep actions grounded:

1. cu-screenshot --base64 (observe)
2. Decide next action(s)
3. Run one action command (or a short chain) and always capture a new screenshot:
   • click/move/type/scroll/drag → screenshot
4. Repeat until done

Commands

All commands print a single JSON object to stdout. Most accept:

• --display :N (defaults to $DISPLAY if set)
• --base64 to include screenshot_base64_png (can be large)
• --out PATH for screenshot output (default: $TWILL_SCREENSHOTS_DIR/...png)
• --no-screenshot to skip the post-action screenshot (use sparingly)

scripts/cu-info

Prints JSON describing the detected screen size, chosen API size, and scale factors.

scripts/cu-screenshot

Takes a screenshot (optionally scaled down to a target like XGA). Use --base64 when your agent needs the image content inline.

scripts/cu-move --x INT --y INT

Moves the mouse pointer.

scripts/cu-click --x INT --y INT

Left-click at a coordinate. Use --button right|middle|left or scripts/cu-right-click.

scripts/cu-double-click, scripts/cu-right-click, scripts/cu-middle-click

Click variants.

scripts/cu-drag --x0 INT --y0 INT --x1 INT --y1 INT

Drag from start to end coordinate.

scripts/cu-type --text STRING

Types text using xdotool type with a small delay.

scripts/cu-key --keys STRING

Sends a key chord (e.g. ctrl+l, Return, Alt+F4) via xdotool key --.

scripts/cu-scroll --dir up|down|left|right --amount INT [--x INT --y INT]

Scrolls by repeated scroll wheel clicks (4/5/6/7). Optionally moves to a coordinate first.

scripts/cu-wait --seconds FLOAT

Sleeps then returns a screenshot.

scripts/cu-cursor

Returns current cursor position (in API space if scaling is enabled).

scripts/cu-zoom --x0 INT --y0 INT --x1 INT --y1 INT

Returns a cropped screenshot of the region (like a “zoom” tool).

Dependencies

Install these in your Ubuntu image:

apt-get update && apt-get install -y --no-install-recommends \
  xdotool \
  scrot \
  imagemagick \
  x11-utils \
  && rm -rf /var/lib/apt/lists/*

If you run a headless desktop in-container, you typically also need:

apt-get update && apt-get install -y --no-install-recommends \
  xvfb \
  mutter \
  x11vnc \
  && rm -rf /var/lib/apt/lists/*

Notes and pitfalls

• Prefer a maximum screenshot size like XGA (1024×768) to reduce image tokens; scaling is automatic when the screen is larger and has a similar aspect ratio.
• If your window manager is missing, GUI apps may not render or focus correctly even though X11 is running.
• If actions seem “offset”, confirm the agent is using API space coordinates from the most recent screenshot and that scaling is enabled.