An autonomous AI Agent that can See, Think, and Act to perform any task on Windows 11 using Ollama with Qwen3-VL vision-language model.

• 👁️ See: Captures screenshots of your desktop using efficient Windows APIs
• 🧠 Think: Analyzes screenshots with Qwen3-VL model via Ollama to understand the UI and plan actions
• 🎯 Act: Executes mouse clicks, keyboard input, and system actions autonomously
• 🔄 Loop: Continues until the task is complete or max iterations reached
• 📸 Screenshot History: Saves all screenshots for debugging and analysis
• 🛡️ Safe Operation: Includes failsafe mechanisms to prevent runaway automation

1. Windows 11 (or Windows 10)
2. Python 3.9+
3. Ollama installed and running
4. Qwen3-VL model pulled in Ollama

1. Install Ollama (if not already installed):

   # Download from https://ollama.ai and install
   # Or use winget:
   winget install Ollama.Ollama

2. Pull the Qwen3-VL model:

   ollama pull qwen3-vl:235b-cloud

3. Clone this repository:

   git clone <repository-url>
   cd Computer-Use-Agent

4. Install Python dependencies:

   pip install -r requirements.txt

python see_think_act_agent.py

You can also run custom tasks programmatically:

from see_think_act_agent import SeeThinkActAgent

# Initialize agent
agent = SeeThinkActAgent(
    model="qwen3-vl:235b-cloud",
    max_iterations=30,
    save_screenshots=True
)

# Run a task
result = agent.run("Open Notepad and type 'Hello World'")
print(result)

Open and run see_think_act_demo.ipynb for interactive examples:

jupyter notebook see_think_act_demo.ipynb

STC/
├── see_think_act_agent.py       # Main agent implementation
├── see_think_act_demo.ipynb     # Demo notebook with examples
├── requirements.txt              # Python dependencies
├── README.md                     # This file
├── utils/
│   ├── __init__.py
│   ├── screenshot_capture.py    # Screen capture utility
│   ├── ollama_client.py         # Ollama API wrapper
│   ├── action_executor.py       # Action execution (mouse/keyboard)
│   └── agent_function_call.py   # Function calling definitions
├── screenshots/                  # Default screenshot directory
└── agent_screenshots/           # Agent execution screenshots

The agent operates in a continuous See-Think-Act loop:

• Captures a screenshot of the current desktop state
• Uses mss library for efficient, low-latency screen capture
• Encodes image for transmission to the model

• Sends screenshot to Qwen3-VL model via Ollama
• Model analyzes the visual state and understands the UI
• Uses function calling to structure the next action
• Considers task progress and decides next step

• Executes the action decided by the model
• Uses pyautogui for precise mouse and keyboard control
• Actions include:
  • Mouse clicks (left, right, double, middle)
  • Mouse movement and dragging
  • Keyboard typing and key presses
  • Scrolling
  • Waiting for UI updates
  • Task termination

• Captures new screenshot to see the result
• Loop continues until task completion or max iterations
• Agent adapts based on what it sees

agent = SeeThinkActAgent(
    model="qwen3-vl:235b-cloud",    # Ollama model name
    max_iterations=30,               # Maximum action loops
    save_screenshots=True,           # Save screenshots for debugging
    screenshot_dir="screenshots",    # Screenshot save directory
    log_level="INFO"                 # Logging verbosity
)

Edit utils/action_executor.py to adjust:

• Mouse movement speed
• Typing speed
• Wait times between actions
• Failsafe settings

# Open an application
agent.run("Open Notepad")

# Type text
agent.run("Open Notepad and type 'Hello World'")

# Use calculator
agent.run("Open Calculator and calculate 123 + 456")

# Web browsing
agent.run("Open Microsoft Edge and search for 'Ollama AI'")

# File management
agent.run("Open File Explorer and create a new folder named 'AI_Projects'")

# Multi-step tasks
agent.run("Open Notepad, type a grocery list, and save it as groceries.txt on Desktop")

• Failsafe: Move mouse to top-left corner to abort (pyautogui feature)
• Max Iterations: Prevents infinite loops
• Keyboard Interrupt: Press Ctrl+C to stop
• Screenshot History: Review what the agent did
• Logging: Full activity logs for debugging

# Pull the model
ollama pull qwen3-vl:235b-cloud

# Verify it's available
ollama list

# Check if Ollama is running
ollama serve

# Or restart Ollama service

• Ensure Python has screen capture permissions
• Try running with administrator privileges
• Check antivirus/security software settings

• Verify screen resolution in agent initialization
• Adjust coordinate scaling if needed
• Check for display scaling settings in Windows

• Adjust pyautogui.PAUSE in action_executor.py
• Modify wait times after actions
• Increase iteration delays

1. Start Simple: Begin with simple tasks and build up complexity
2. Clear Desktop: Minimize visual clutter for better recognition
3. Specific Instructions: Give clear, specific task descriptions
4. Monitor Progress: Watch the agent work to understand its decisions
5. Review Screenshots: Check saved screenshots to debug issues
6. Reasonable Scope: Keep tasks focused and achievable

Test individual components:

# Test screenshot capture
python utils/screenshot_capture.py

# Test Ollama client
python utils/ollama_client.py

# Test action executor
python utils/action_executor.py

┌─────────────────────────────────────────────────────────────┐
│                    See-Think-Act Agent                      │
└─────────────────────────────────────────────────────────────┘
                            │
        ┌───────────────────┼───────────────────┐
        │                   │                   │
        ▼                   ▼                   ▼
┌──────────────┐    ┌──────────────┐    ┌──────────────┐
│  SEE Module  │    │ THINK Module │    │  ACT Module  │
│              │    │              │    │              │
│  Screenshot  │───▶│    Ollama    │───▶│  PyAutoGUI   │
│   Capture    │    │   Qwen3-VL   │    │   Control    │
│   (mss)      │    │  Vision LLM  │    │              │
└──────────────┘    └──────────────┘    └──────────────┘
        │                   │                   │
        │                   │                   │
        └───────────────────┴───────────────────┘
                            │
                      ┌─────▼─────┐
                      │  Windows  │
                      │    OS     │
                      └───────────┘

Contributions are welcome! Areas for improvement:

• Additional action types
• Better error handling
• Multi-monitor support
• Task planning and optimization
• Integration with more models

• Qwen Team: For the amazing Qwen3-VL model
• Ollama: For making local LLM inference easy
• PyAutoGUI: For GUI automation capabilities
• MSS: For efficient screen capture

For issues and questions:

• Open an issue on GitHub
• Check existing issues for solutions
• Review the troubleshooting section

⚠️ Warning: This agent has the ability to control your computer. Use responsibly and monitor its actions, especially when testing new tasks.