AGI Agent is an open-source Vibe platform designed for interactive document creation, coding, and general task execution, offering a locally deployable alternative to platforms like Claude Cowork. It targets users seeking powerful local AI agents, professional document generation capabilities, and a Vibe workflow experience, providing a flexible and customizable environment for complex productivity tasks.

How It Works

AGI Agent employs a plan-driven ReAct model with multi-round iterative execution, allowing agents to call tools and process feedback. Its core architecture features an autonomous multi-agent system (Manager + sub-Agents), a ReAct execution engine, and a dual-layer memory system with progressive history compression for long-range tasks. A novel aspect is its built-in semantic and keyword retrieval for long-term memory, eliminating the need for external embedding models. The platform supports a three-tier tool ecosystem (built-in, OS, MCP) and emphasizes interactive creation through a real-time streaming GUI, enabling direct editing of SVG and Mermaid diagrams.

Quick Start & Requirements

• Primary Install: Recommended install.sh script for one-click installation. Alternatively, pip install -r requirements.txt.
• Prerequisites: Python 3.8+ (note: fastmcp requires commenting out for Python 3.8). Optional dependencies include playwright (for Mermaid/web search), pandoc (for Word export), and xelatex/pandoc (for PDF/LaTeX).
• Runtime Interfaces: Terminal, Python Library, and Web Interface modes are supported.
• Web GUI: Run via python GUI/app.py --port 5001 (experimental, single-user dev version).
• Links: Online Demo: https://agiagentonline.com. Demo videos and detailed guides are linked within the README.

Highlighted Details

• Interactive Document Generation: Supports rich image and text editing, with direct SVG and Mermaid diagram manipulation. Offers lossless export to Markdown, Word, PDF, and LaTeX.
• Autonomous Multi-Agent Architecture: A Manager orchestrates multiple sub-Agents (e.g., coder, artist) that can self-organize, communicate, and collaborate.
• Broad LLM Compatibility: Integrates with various LLMs via Anthropic/OpenAI API, including Claude, GPT-4, DeepSeek V3, Kimi K2, GLM, and Qwen, supporting both domestic Chinese and international models.
• Advanced Memory Management: Features progressive history compression and built-in vector retrieval for long-term memory, enabling extended task execution without external dependencies.

Maintenance & Community

The project can be contacted via email (bitcursor@2925.com) or the WeChat public account "AGI智创机器人". Issues and suggestions can be submitted through GitHub Issues. Recent news indicates ongoing development and availability of online services and installers.

Licensing & Compatibility

The project is described as "100% open source" and "completely free," but a specific license type (e.g., MIT, Apache) is not explicitly stated in the provided text, which requires clarification for commercial use or closed-source linking. It is compatible with Linux, Windows, and macOS, and supports local, cloud, or embedded (ARM) deployment.

Limitations & Caveats

The Web GUI is noted as experimental and not suitable for industrial deployment. Users are cautioned about the security implications of the agent's ability to execute system terminal commands, recommending command review, sandbox execution, and regular data backups. Continuing execution via the -c flag does not restore the LLM's conversational context.