How Computer Use Works

Computer Use operates through a continuous feedback loop between Claude and your computer. Understanding this loop is essential for building effective agents.

The Agentic Loop

The loop has exactly one dangerous step, and it is yours. Steps 1-3 happen inside the API. Executing the action runs on your machine, with your permissions. Every safety control in Module 5 attaches to that one step, because it is the only one you can intercept.

The Computer Tool

The computer_20251124 tool version provides these actions:

Action	Description
screenshot	Capture current screen state
mouse_move	Move cursor to x,y coordinates
left_click	Click left mouse button
right_click	Click right mouse button
double_click	Double-click left button
triple_click	Select entire line/paragraph
left_mouse_down	Hold left button (for dragging)
left_mouse_up	Release left button
scroll	Scroll up/down/left/right
type	Type text string
key	Press keyboard key or combo
hold_key	Hold a key while performing action
wait	Pause for specified duration
zoom	View a specific screen region at full resolution (requires enable_zoom: true)

Required Headers

Computer Use is opt-in per request. Include the beta header:

headers = {
    "anthropic-beta": "computer-use-2025-11-24"
}

Screen Resolution

Claude works best with specific resolutions. The recommended setup:

# Optimal resolutions for Computer Use
RECOMMENDED_RESOLUTIONS = [
    (1024, 768),   # XGA - fast, efficient
    (1280, 800),   # WXGA - good balance
    (1920, 1080),  # Full HD - more detail
]

Tip: Lower resolutions mean faster screenshot processing and lower token costs.

Vision-Based Understanding

Claude uses its vision capabilities to:

1. Identify UI elements (buttons, text fields, menus)
2. Read text on screen
3. Understand layout and spatial relationships
4. Track changes between screenshots

This means Claude can work with any application without needing specialized integrations.

Next, we'll look at the Agent SDK that simplifies building these loops. :::