autogen-multi-agent-chat

# AutoGen Multi-Agent Chat

An industrial-grade, multi-agent AI coding platform built on [Microsoft AutoGen](https://github.com/ag2ai/ag2). Three specialized AI agents collaborate autonomously to analyze requirements, write code, review it, and execute — featuring a full-stack React + FastAPI web interface.

## Features

- **Three-Agent Architecture** — Admin (executor), Architect (coder), Reviewer (auditor)
- **Local Safe Sandbox** — Regex-based guardrails block dangerous commands; subprocess timeout prevents infinite loops
- **Agent Toolchain** — File read/write, web search (DuckDuckGo), log grep — with project-directory isolation
- **ChromaDB Memory** — Long-context compression via vector store; recalls relevant past context across sessions
- **Full-Stack Web UI** — React + TypeScript + Xterm.js frontend with FastAPI WebSocket backend
- **CLI Interface** — `python main.py -t "your task"` for terminal-based usage
- **Cost-Effective Mode** — Use high-tier models for coding, low-tier for review/routing

## Architecture

```
┌─────────────────────────────────────────────────────────────────┐
│                        FastAPI Server                           │
│  REST: /api/files, /api/models    WebSocket: /ws (streaming)   │
├────────────────────────────────────────────────────────┬────────┤
│                 AutoGen GroupChat                      │ Memory │
│  ┌─────────┐  ┌────────────┐  ┌──────────┐           │ Chrome │
│  │  Admin   │  │  Architect │  │ Reviewer │           │   DB   │
│  │ executor │  │   coder    │  │  auditor │           │        │
│  └────┬─────┘  └─────┬──────┘  └──────────┘           │        │
│       │              │                                 │        │
│  ┌────▼──────────────▼─────┐   ┌──────────────────┐   │        │
│  │   LocalSafeWorkspace    │   │    Toolchain      │   │        │
│  │  • danger regex guard   │   │  • read/write     │   │        │
│  │  • 30s timeout          │   │  • web search     │   │        │
│  │  • coding/ sandbox      │   │  • grep logs      │   │        │
│  └─────────────────────────┘   └──────────────────┘   │        │
├───────────────────────────────────────────────────────┴────────┤
│                    React + Xterm.js Frontend                    │
│   ┌──────────┬──────────────────┬─────────────────────┐        │
│   │File Tree │   Agent Chat     │ Code Viewer + Term  │        │
│   └──────────┴──────────────────┴─────────────────────┘        │
└─────────────────────────────────────────────────────────────────┘
```

## Project Structure

```
├── main.py                  # CLI entry point
├── server.py                # FastAPI + WebSocket backend
├── start.sh                 # One-click launcher (dev & prod)
├── src/
│   ├── config.py            # Env loading, tiered model configs
│   ├── agents.py            # Agent factories & system prompts
│   ├── chat_manager.py      # GroupChat orchestration + memory hooks
│   ├── sandbox.py           # LocalSafeWorkspace with guardrails
│   ├── tools.py             # File I/O, web search, grep tools
│   ├── memory.py            # ChromaDB vector memory manager
│   └── utils.py             # Coloured logging
├── frontend/
│   ├── src/App.tsx           # React SPA (chat + file tree + terminal)
│   ├── src/index.css         # Dark theme styles
│   └── vite.config.ts        # Vite config with API proxy
├── coding/                   # Execution sandbox
├── requirements.txt
└── .env.example
```

## Quick Start

### One-Click Launch

```bash
git clone https://github.com/Ei-Ayw/autogen-multi-agent-chat.git
cd autogen-multi-agent-chat
cp .env.example .env   # Edit with your API key

# Production (builds frontend, serves at :8000)
./start.sh

# Development (Vite HMR at :5173, FastAPI at :8000)
./start.sh --dev
```

### CLI Mode

```bash
python main.py -t "Write a snake game in Python"
python main.py -f requirements.md
python main.py -t "Build a REST API" --cost-effective
python main.py --help
```

### Manual Setup

```bash
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Frontend
cd frontend && npm install && npm run build && cd ..

# Run
uvicorn server:app --host 0.0.0.0 --port 8000
```

## Configuration

All settings via `.env`:

| Variable | Description | Default |
|----------|-------------|---------|
| `API_TYPE` | `openai` (relay) or `anthropic` (native SDK) | `openai` |
| `MODEL` | Default model | `claude-opus-4-6` |
| `ANTHROPIC_API_KEY` | API key | — |
| `ANTHROPIC_BASE_URL` | API endpoint / relay URL | — |
| `SANDBOX_TIMEOUT` | Max code execution time (seconds) | `30` |

### Cost-Effective Mode

| Agent | Model | Purpose |
|-------|-------|---------|
| Architect | `claude-opus-4-6` | Heavy reasoning & code generation |
| Reviewer  | `claude-haiku-4-5-20251001` | Fast, cheap code review |
| Manager   | `claude-haiku-4-5-20251001` | Speaker selection routing |

## Security

- **Sandbox guardrails**: Regex patterns block `rm -rf`, `mkfs`, fork bombs, `shutil.rmtree('/')`, etc.
- **Execution timeout**: Subprocess killed after 30s (configurable)
- **Path isolation**: File tools restricted to project directory; path traversal blocked
- **No Docker required**: Safe local execution without container overhead

## Requirements

- Python 3.8+
- Node.js 18+ (for frontend)
- An accessible LLM API endpoint

## License

[MIT](LICENSE)

---

## 中文说明

对标 Devin 的工业级多智能体 AI 编程平台。三个 Agent 自主协作：分析需求 → 编写代码 → 审查 → 执行，配备安全沙箱、向量记忆、全栈 Web UI。

```bash
git clone https://github.com/Ei-Ayw/autogen-multi-agent-chat.git
cd autogen-multi-agent-chat
cp .env.example .env     # 填入你的 API Key
./start.sh               # 一键启动，访问 http://localhost:8000
```

CLI 模式：`python main.py -t "写一个贪吃蛇游戏"`