Magentic-UI is a research prototype for a human-centered web agent, designed for automating complex web tasks that require user interaction and control. It targets users needing to perform intricate web actions, navigate deep site structures, or combine web browsing with code execution, offering a transparent interface for human-in-the-loop collaboration.

How It Works

Magentic-UI employs a multi-agent system based on AutoGen's Magentic-One. Key agents include an Orchestrator (LLM-powered, manages planning and delegation), WebSurfer (LLM with browser control for advanced web interaction), Coder (LLM with Docker for code execution), and FileSurfer (LLM for file operations). This modular design allows for collaborative planning, interruptible task execution, action guards for sensitive operations, and plan learning/retrieval.

Quick Start & Requirements

• Install: pip install magentic-ui
• Run: magentic ui --port 8081
• Prerequisites: Python 3.10+, Docker, WSL2 (for Windows). Requires an OpenAI API key (configurable for Azure/Ollama via config.yaml or UI settings).
• Setup: Initial Docker image build can take time; subsequent runs are faster.
• Docs: How to Use

Highlighted Details

• Co-Planning & Co-Tasking: Enables collaborative step-by-step plan creation and allows users to interrupt and guide task execution.
• Action Guards: Ensures explicit user approval for sensitive actions.
• Advanced WebSurfer: Improved browser control over AutoGen's MultimodalWebSurfer, supporting tab management, file uploads, and multimodal queries.
• Plan Gallery: Learns from past runs and allows retrieval of saved plans for future automation.

Maintenance & Community

This is a Microsoft research project. Contributions are welcomed via issues and pull requests, with specific labels for community involvement. The project follows the Microsoft Open Source Code of Conduct.

Licensing & Compatibility

Licensed under the MIT License. Use of Microsoft trademarks is subject to their guidelines.

Limitations & Caveats

Magentic-UI is a research prototype, implying potential for instability or breaking changes. Building from source requires Node.js and frontend build steps.