What the agent can do

Models

Every listed model supports vision and tool calling so the agent can see the desktop and drive mouse, keyboard and shell tools. Claude Opus 5.5 is the default for computer use; Claude Sonnet 5 is the fast option. GPT-6.1 Sol, Gemini 4 Argon and Kimi K3 (open weights) are also available.

Vision

The agent can see its desktop by taking screenshots. Before any action, it typically takes a screenshot to understand the current state.

• Captures the full 1280×720 desktop
• Identifies UI elements, text and positions
• Used before and after actions to verify results

Mouse control

• Click: Left, right or double-click at any position
• Move: Position the cursor
• Scroll: Scroll up or down

Keyboard control

• Type: Enter text in input fields
• Press: Single keys (Enter, Escape, Tab, etc.)
• Hotkeys: Key combinations (Ctrl+C, Alt+Tab, etc.)

Shell commands

The agent can execute any shell command in the sandbox:

• Run scripts and programs
• Install packages (apt-get)
• Manage files and directories
• Check system information
• Start background processes

File operations

• Read: View contents of any file
• Write: Create or modify files

Pre-installed software

Each sandbox is a Debian 13 desktop with:

• Chromium: Dark-themed web browser
• Thunar: File manager
• XFCE Terminal: Terminal emulator
• Python 3: Scripting language
• curl, wget: HTTP clients
• git: Version control
• vim, nano: Text editors

Limitations