Computer Control

Computer Control Skill

Use Claude's Computer Use API to see and control desktop environments through screenshots and mouse/keyboard actions.

When To Use

• Automating GUI-based workflows that lack CLI alternatives
• Testing web applications through visual interaction
• Filling forms, navigating menus, or interacting with desktop apps
• Building automation pipelines that need visual verification

When NOT To Use

• Tasks achievable through CLI or API (no GUI needed)
• Browser automation better served by Playwright or CDP

Why this stays opt-in. Per docs/inclusive-defaults.md (TRUE-exception category 4), Computer Use takes screenshots and synthesizes keyboard/mouse input: cross-process side effects that must always be explicitly invoked, never default-on.

Architecture

The computer use system has three layers:

1. Display Toolkit (phantom.display) - executes OS-level actions via xdotool/scrot on the real or virtual display
2. Agent Loop (phantom.loop) - manages the conversation cycle between Claude API and the display toolkit
3. CLI (phantom.cli) - command-line interface for running tasks or checking environment readiness

User Task
    |
    v
Agent Loop  <---->  Claude API (beta)
    |                   |
    v                   v
Display Toolkit    tool_use responses
    |              (click, type, screenshot)
    v
OS Commands (xdotool, scrot)
    |
    v
Display (X11 / Xvfb / WSLg)

Quick Start

Check environment

cd plugins/phantom
uv run python -m phantom.cli --check

Run a task

export ANTHROPIC_API_KEY="sk-ant-..."
uv run python -m phantom.cli "Open Firefox and search for Claude AI"

Use in Python

from phantom.display import DisplayConfig, DisplayToolkit
from phantom.loop import LoopConfig, run_loop

result = run_loop(
    task="Take a screenshot of the desktop",
    api_key="sk-ant-...",
    loop_config=LoopConfig(
        model="claude-sonnet-5",
        max_iterations=10,
    ),
    display_config=DisplayConfig(width=1920, height=1080),
)

print(f"Done in {result.iterations} iterations")
print(result.final_text)

API Versions

The resolve_tool_version() function handles this mapping automatically based on the model name.

Available Actions

All versions:

• screenshot - capture display
• left_click - click at [x, y]
• type - type text string
• key - press key combo (e.g., ctrl+s)
• mouse_move - move cursor

Enhanced (20250124+):

• scroll - scroll with direction and amount
• left_click_drag - drag between coordinates
• right_click, middle_click, double_click, triple_click
• hold_key - hold key for duration
• wait - pause between actions

Latest (20251124):

• zoom - inspect screen region at full resolution

Safety

Computer use carries risks. Follow these guidelines:

1. Use a sandbox: Run in Docker or a VM, not your main OS
2. Limit access: Do not provide login credentials unless necessary, and never for banking or sensitive services
3. Set iteration caps: Always use max_iterations to prevent runaway API costs
4. Human approval: For actions with real-world consequences, add confirmation callbacks via on_action
5. Close sensitive apps: Claude sees the full screen via screenshots; close anything private before starting

Environment Requirements

Linux (native or WSL2 with WSLg):

sudo apt install xdotool scrot xclip

Headless (Docker/CI):

# Install Xvfb for virtual display
sudo apt install xvfb xdotool scrot xclip
Xvfb :1 -screen 0 1920x1080x24 &
export DISPLAY=:1

Prompting Tips

1. Be specific about each step of the task
2. Add "After each step, take a screenshot and verify" to catch mistakes early
3. Use keyboard shortcuts when UI elements are hard to click
4. Provide example screenshots for repeatable workflows
5. Set a system prompt with domain-specific instructions

Exit Criteria

• uv run python -m phantom.cli --check exits 0 before any task is launched; if it fails, required OS tools (xdotool, scrot, xclip) are installed or Xvfb is started before proceeding
• max_iterations set on every run_loop() call; no task launched without an explicit iteration cap to prevent runaway API costs
• Final result includes result.iterations count and result.final_text confirming the task outcome; empty final_text treated as failure, not success
• Sensitive applications (password managers, banking, private files) closed before task starts; task prompt does not contain raw credentials