Xiajiao IMAI Agent Team Collaboration

6 deps, npm start — manage your AI agents like managing a team

Quick Start →

Recipes

GitHub

🤖

Multi-Agent Chat

Create groups, add agents, @mention for precise routing. Agents talk to each other and collaborate like a real team.

Learn more

🔧

Tool Calling

7 built-in tools + extensible custom tools — HTTP tools (zero-code), JS auto-register, and MCP bridged. Web search, RAG, memory, cross-agent calls, channels, schedules, and any REST API.

Learn more

🧠

Persistent Memory

Three-category memory (semantic / episodic / procedural), embedding dedup. The more you use an agent, the better it knows you.

Learn more

📚

RAG Knowledge Base

BM25 + vector hybrid retrieval + RRF + LLM re-ranking. Upload docs, agents auto-index and auto-retrieve.

Learn more

⚡

Minimal Setup

6 npm deps, zero external services. No Docker, PostgreSQL, or Redis needed — just npm start.

Deploy guide

🔗

Collaboration Flow

Auto-relay collaboration chains + real-time visual panel + human intervention. One message triggers the entire agent pipeline.

Learn more

💬 Drive Your Agent Team Through Chat

Chat with agents via IM — they auto-invoke tools, query status, manage channels — all visible in real time.

Xiajiao IM — Collaboration Flow Demo

npm deps

external services

built-in + custom tools

unit tests

🚀 Up and Running in 30 Seconds

bash

git clone https://github.com/moziio/xiajiao.git
cd xiajiao && npm install   # 6 deps, done in seconds
npm start                    # Open http://localhost:18800

No Docker, no PostgreSQL / Redis (built-in SQLite), no extra env vars. Just these three lines.

💡 What It Feels Like

🎭 Scenario 1: AI Writing Studio

Group: Novelist + Editor + Translator | Chain: auto-relay

You: Write an essay about indie developers
→ Novelist writes 800-word Chinese essay
→ Editor auto-relays to polish
→ Translator auto-translates to English
→ 2 min, bilingual masterpiece ready

📖 Scenario 2: Private Knowledge Assistant

Agent tools: RAG + Persistent Memory

You: Upload API docs to knowledge base
You: How do I call the payment API?
→ Agent retrieves from YOUR docs → accurate answer, no hallucination

⚔️ Scenario 3: Multi-Model Arena

3 agents, same SOUL.md, different models

You: @GPT @Claude @Qwen Explain attention mechanism
→ Three models answer side by side
→ Compare quality, depth, style at a glance

See all 12 recipes →

📸 Real Screenshots

These are from a live Xiajiao instance, not mockups.

Collaboration chain + AI-generated summer night art

Collaboration chain + AI art generation

Group chat @mention routing

Poetry + auto-translation relay

Agent poetry + AI image generation

News Agent — search & structured output

🔬 Why So Lightweight?

Where others use frameworks, Xiajiao uses Node.js standard libraries. Where others need external services, Xiajiao uses SQLite.

🚫 No Express

node:http handles 15 endpoints. One less dep = one less security risk.

🚫 No PostgreSQL

node:sqlite built-in DB. WAL for concurrent reads, FTS5 for full-text search.

🚫 No React/Vue

Vanilla JS frontend. Edit code, refresh browser — no Webpack build step.

🚫 No Jest

node:test runs 53 tests. Zero test deps, zero config files.

See full architecture →

🏗️ How It Compares

Dimension	Xiajiao	Dify / FastGPT
Install	`npm start`	`docker compose up` (needs PostgreSQL + Redis)
Deps	6 npm packages	Multiple packages + external services
Agent UX	IM group chat + @mention	Workflow canvas
Agent Relations	Equal collaboration, mutual @mention	Predefined DAG pipelines
Data Storage	Fully local SQLite	Requires DB config
Philosophy	Agents are your colleagues	Agents are your apps

📝 SOUL.md — Define Agents with Markdown

markdown

# Translator

You are a bilingual translation expert (Chinese ↔ English).

## Principles
- Faithfulness, expressiveness, elegance
- Output translation directly, no word-by-word analysis

## Restrictions
- Do not translate content inside code blocks
- Do not proactively @mention other agents

Share a single .md file to clone an agent's personality. Git version control, diffs at a glance.

SOUL.md writing guide →

🔌 All Major Models Supported

Provider	Models	Highlights
OpenAI	GPT-4o / o1	Most comprehensive
Anthropic	Claude Sonnet / Opus	Code king
Qwen	Qwen Max / Turbo	Chinese optimized, low cost
DeepSeek	Chat / Coder / Reasoner	Best value
Ollama	Llama 3 / Qwen 2	Completely free, runs locally
OpenRouter	100+ model aggregator	One key for all

Different agents can use different models — Claude for coding, GPT-4o for translation, Qwen for daily chat.

Full model config guide →

❓ Common Questions

How is Xiajiao different from ChatGPT / Claude?

ChatGPT is "one AI chatting". Xiajiao is "a team of AIs collaborating" — you create multiple agents, each with their own personality and skills, they @mention each other in groups and auto-relay. All data stays on your machine.

Is it free?

Xiajiao is fully open source (MIT). You only pay for LLM API calls — use Ollama local models and even that's free.

How does it compare to Dify?

Dify is an "AI app development platform" — for building AI products for end users. Xiajiao is "agent team collaboration" — agents are your colleagues. Detailed comparison →

Is it secure? Where does my data go?

All data stored locally. Zero telemetry. The only external communication is the LLM API you configure. Use Ollama = fully offline. Security details →

See 30+ FAQs →

Xiajiao (虾饺) — named after the Cantonese shrimp dumpling: small, delicate, packed with flavor.

What is Xiajiao · Quick Start · FAQ · Changelog · GitHub