Xiajiao

Xiajiao (虾饺) — named after the Cantonese shrimp dumpling: small, delicate, but packed with flavor.
Minimal dependencies, maximum capability.

⚡ Quick Start

Option 1: npm (recommended)

Prerequisite: Node.js >= 22.0.0

git clone https://github.com/moziio/xiajiao.git
cd xiajiao && npm install
npm start

Option 2: Docker

git clone https://github.com/moziio/xiajiao.git
cd xiajiao
docker build -t xiajiao .
docker run -d -p 18800:18800 \
  -v xiajiao-data:/app/data \
  -v xiajiao-uploads:/app/public/uploads \
  --name xiajiao xiajiao

Open http://localhost:18800 → Log in with default password admin → Settings → Models to add your API key → Start chatting!

[!TIP] Password and API keys can all be changed in the web UI — no config files to edit.

🤔 Why Xiajiao?

Most AI platforms require Docker + PostgreSQL + Redis + vector databases just to get started.
Xiajiao takes a radically different approach — it's the only IM platform where your agents actually collaborate as a team:

Who is this for?

• Solo developers who want an AI team without the DevOps overhead
• AI enthusiasts exploring multi-agent collaboration patterns
• Small teams needing a self-hosted AI workspace with zero vendor lock-in
• Researchers prototyping agent-to-agent communication and memory systems

The core difference: Dify and Coze are AI app development platforms. Xiajiao is an AI Agent team collaboration platform — treat agents as teammates, not tools.

✨ Features

🤖 Multi-Agent Group Chat

Create groups, add agents, use @mentions to route messages. Agents can talk to each other, hand off tasks, and collaborate — just like a real team.

🔧 Tool Calling

Agents don't just chat — they take action. 7 built-in tools with an extensible architecture:

🧠 Persistent Agent Memory

A 3-class memory system that makes agents smarter over time:

• Semantic memory — facts & knowledge ("User prefers Python")
• Episodic memory — conversation events ("We discussed deployment last time")
• Procedural memory — behavioral patterns ("Keep replies concise")

Embedding-based dedup + hybrid search + automatic prompt injection.

📚 Production-Grade RAG

Upload documents — agents auto-index and auto-retrieve:

• BM25 + vector hybrid retrieval
• Reciprocal Rank Fusion (RRF)
• LLM reranking
• Hierarchical chunking (200-char small chunks + 800-char large chunks)

🔗 Collaboration Flows

• Collaboration chains — agents relay within a group; one's output becomes the next's input
• Visual flow panel — real-time status, history replay, human intervention (approve / stop / edit)
• Workflow engine — multi-step, conditional branching, error handling (retry / skip / rollback), human approval

🔌 Multi-Model Support

Works with any OpenAI-compatible API provider:

OpenAI (GPT-4o) · Anthropic (Claude) · Qwen (通义) · GLM · Kimi · MiniMax · DeepSeek · Ollama

📋 And More

🏗️ Tech Stack

Why so few? We treat every dependency as a liability, not a feature. If the standard library can do it, we don't add a package.

⚙️ Configuration

Environment Variables

Model Configuration

cp models.example.json models.json

{
  "providers": {
    "openai": {
      "baseUrl": "https://api.openai.com/v1",
      "apiKey": "sk-your-api-key",
      "api": "openai-completions"
    }
  },
  "models": [
    {
      "id": "openai/gpt-4o",
      "name": "gpt-4o",
      "provider": "openai",
      "reasoning": false,
      "input": ["text", "image"],
      "contextWindow": 128000,
      "maxTokens": 4096
    }
  ]
}

Supported API types: openai-completions (Qwen / Kimi / GLM / DeepSeek / Ollama) · anthropic-messages · dashscope-image

[!TIP] You can also add models directly in the web UI under Settings → Models — no need to edit files manually.

📁 Project Structure

xiajiao/
├── server/                   # Backend
│   ├── index.js              # HTTP + WebSocket entry
│   ├── middleware/            # Auth, logging, rate limiting
│   ├── routes/               # API routes
│   ├── services/             # Core services
│   │   ├── llm.js            # LLM + Tool Calling loop
│   │   ├── rag.js            # RAG retrieval engine
│   │   ├── memory.js         # Agent memory
│   │   ├── workflow.js       # Workflow engine
│   │   ├── collab-flow.js    # Collaboration flow state machine
│   │   ├── tool-registry.js  # Tool registry
│   │   └── database.js       # SQLite + migrations
│   ├── migrations/           # Database migrations
│   └── tests/                # Unit tests
├── public/                   # Frontend (Vanilla JS, no build step)
│   ├── js/                   # Feature modules
│   └── css/
├── data/_soul-templates/     # Default agent SOUL.md templates
├── models.example.json       # Model config example
├── im-settings.example.json  # System settings example
└── agents.example.json       # Agent config example (5 built-in agents)

🗺️ Roadmap

🚀 Xiajiao is under active development. Here's what's shipped and what's coming:

This is just the beginning. Got ideas? Open an Issue or join the discussion.

🤝 Contributing

We welcome all contributions — bug reports, feature ideas, code, docs, and translations.

See CONTRIBUTING.md for details.

⭐ Show Your Support

[!IMPORTANT] Star this repo — you'll receive notifications for every new release, and it helps others discover the project!

🔒 Security

To protect users, please do not post security issues publicly on GitHub Issues. Instead, please email security concerns to the maintainers directly. We will respond promptly.

📄 License

MIT