--- name: computer-use description: Control a virtual desktop (X11 via Xvfb) using Anthropic Claude's computer-use tool API. Can open applications, click, type, screenshot, drag — simulating a real user on a GUI. Use when a task requires real desktop interaction (not just browser automation) or visual UI navigation. ---
Computer Use Skill (Anthropic Claude)
Gives Claude a virtual screen it can see and control. Lightweight setup: Xvfb + xdotool + scrot + Chromium (already installed).
Start the virtual desktop
bash
bash /root/.openclaw/workspace/skills/computer-use/start_desktop.sh
Creates Xvfb :99 + Chromium window at 1280x800
Run an agent task
bash
python3 /root/.openclaw/workspace/skills/computer-use/agent.py --task "Open Wikipedia and search for civil engineering"
Stop the virtual desktop
bash
bash /root/.openclaw/workspace/skills/computer-use/stop_desktop.sh
Architecture
- Xvfb :99 → virtual framebuffer (no GPU needed)
- xdotool → simulate keyboard/mouse
- scrot → take screenshots
- Chromium → browser inside the VFB
- Claude computer-use API → vision model that decides actions
Costs
- Uses ANTHROPIC_API_KEY
- Claude Sonnet 4.5 pricing: $3/M input, $15/M output
- Typical task (10 screenshots + actions): ~$0.10-0.50