中文 | English

AutoGUI

AI-driven screen automation MCP Server. Send natural language tasks, and the internal AI captures screenshots, analyzes them, and performs mouse/keyboard actions autonomously.

Architecture

MCP Client (Claude Code, Claude Desktop, Cursor, etc.)
    |  stdio
    v
server.py (FastMCP async orchestration loop)
    |
    v
agent.py (ScreenAgent toolkit: capture, execute, parse, safety)

Quick Start

git clone https://github.com/stellariums/AutoGUI.git
cd AutoGUI
pip install -r requirements.txt

Configuration

AutoGUI supports layered configuration: environment variables (highest priority) > config.json > defaults.

Option A: Environment Variables Only (Simplest)

Only 3 variables needed to get started:

set AUTOGUI_API_KEY=your-api-key
set AUTOGUI_BASE_URL=https://api.openai.com/v1
set AUTOGUI_MODEL=gpt-4o

Or copy .env.example to .env and pass via your MCP client config (see below).

Option B: Config File

cp config.json.example config.json

Edit config.json — only the api section is required, everything else has sensible defaults:

{
  "api": {
    "base_url": "https://your-api-endpoint/v1",
    "api_key": "your-api-key",
    "model": "your-model-name"
  }
}

{
  "screen": {
    "max_width": 1280,
    "max_height": 720,
    "allowed_region": null
  },
  "agent": {
    "max_iterations": 20,
    "delay_between_actions": 1.0,
    "max_history_rounds": 10
  },
  "safety": {
    "enable_confirmation": true,
    "fallback_action": "block",
    "dangerous_keys": ["delete", "backspace", "escape"],
    "dangerous_hotkeys": [["ctrl", "w"], ["alt", "f4"]],
    "dangerous_patterns": ["rm ", "del ", "format ", "shutdown"]
  }
}

MCP Client Setup

Claude Code

claude mcp add AutoGUI -- python /path/to/AutoGUI/server.py

Or add to .mcp.json:

{
  "mcpServers": {
    "AutoGUI": {
      "command": "python",
      "args": ["/path/to/AutoGUI/server.py"],
      "env": {
        "AUTOGUI_API_KEY": "your-api-key",
        "AUTOGUI_BASE_URL": "https://api.openai.com/v1",
        "AUTOGUI_MODEL": "gpt-4o"
      }
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "AutoGUI": {
      "command": "python",
      "args": ["C:/path/to/AutoGUI/server.py"],
      "env": {
        "AUTOGUI_API_KEY": "your-api-key",
        "AUTOGUI_BASE_URL": "https://api.openai.com/v1",
        "AUTOGUI_MODEL": "gpt-4o"
      }
    }
  }
}

Cursor

Add to Cursor MCP settings (.cursor/mcp.json):

{
  "mcpServers": {
    "AutoGUI": {
      "command": "python",
      "args": ["/path/to/AutoGUI/server.py"],
      "env": {
        "AUTOGUI_API_KEY": "your-api-key",
        "AUTOGUI_BASE_URL": "https://api.openai.com/v1",
        "AUTOGUI_MODEL": "gpt-4o"
      }
    }
  }
}

MCP Inspector (Testing)

npx @modelcontextprotocol/inspector python server.py

Tool

Supported Actions

Safety

• Dangerous action detection (rule-based + AI self-labeling)
• Configurable dangerous keys, hotkeys, and text patterns
• Optional region restriction (allowed_region)
• Elicit-based confirmation for dangerous operations
• Configurable fallback: block or allow

FAQ

Q: Screenshot is black or empty A: Make sure the screen is not locked. On Windows, pyautogui/mss cannot capture the lock screen.

Q: Chinese input not working A: AutoGUI uses clipboard (pyperclip + Ctrl+V) for text input, which supports CJK characters. Make sure pyperclip is installed.

Q: "API key required" error A: Set AUTOGUI_API_KEY env var or add api.api_key in config.json.

Q: "Another task is already running" error A: AutoGUI processes one task at a time. Wait for the current task to finish.

Requirements

• Windows 10/11
• Python >= 3.10
• An OpenAI-compatible vision API (GPT-4o, Qwen-VL, etc.)

Acknowledgments

This project is based on tech-shrimp/qwen_autogui, refactored as an MCP Server with added safety checks, layered configuration, and multi-client support.

License

MIT