中文文档 · GitHub

MChat is a lightweight, embeddable, multi-tenant vertical RAG platform. It combines a streaming Bot engine, production-grade RAG knowledge base, hot-reload Skill plugins, visual Workflow orchestration (Beta), and a one-line embeddable Widget — with 10+ LLM providers and multi-channel delivery (Web Widget, WebSocket, REST, WeChat Official Account, and more).

Built-in AI customer service works out of the box. Extend the same stack into vertical channels — patent search, medical, legal, and other domain RAG packages with dedicated knowledge bases, skill packs, and tuned retrieval — embed anywhere with a single <script> tag.

• English main site
• Chinese main site
• Full screenshot tour
• Product roadmap
• Workflow orchestrator (Beta)

Click any screenshot to open the full image.

• Bot engine — Streaming LLM inference + tool calling; OpenAI, Anthropic, Google, DeepSeek, Ollama, Groq, and more
• Skill plugins — Hot-reload SKILL.md from disk/zip/URL, OpenClaw-compatible formats; premium packs as vertical add-ons
• RAG knowledge base — Multi-strategy chunking, multi-provider embeddings, hybrid retrieval (vector + BM25 + RRF), rerank, query rewriting, parent-child context
• Workflow orchestration (Beta) — React Flow DAG editor, linear or graph execution, merge/condition/approval nodes, built-in report templates, manual/schedule/channel triggers
• Embeddable Widget — One <script> tag for branded vertical RAG chat on any website
• Multi-tenant — Independent channel configs with isolated AI, skills, and knowledge bases
• Vertical channels — One-click domain templates: model, prompt, KB, rerank, skills, widget theme
• Multi-channel — Web Widget, REST, WebSocket, WeChat (DingTalk/WhatsApp/Telegram planned)
• Speech input — OpenAI Whisper (optional local models)
• Security — JWT, API keys, RBAC
• Docker — make docker-up-lite full stack

MChat Workflow chains multiple Skills into reusable pipelines — not just one-shot chat, but multi-step automation with parallel branches, conditional routing, human approval, and structured outputs (reports, exports, alerts).

flowchart LR
  T[Manual / Schedule / Channel] --> W[Workflow DAG]
  W --> S1[Skill nodes]
  W --> S2[Skill nodes]
  S1 --> M[Merge]
  S2 --> M
  M --> E[End / Export / Notify]

Visual graph editor (Admin → Workflows, /admin/workflows):

• ComfyUI-style canvas: pointer vs pan (V / H), drag nodes from the skill library
• Node types: start · skill · condition · approval · merge · end
• Payload mapper — ${input.keyword}, ${nodes.<id>.result.xxx} templates per node
• Merge — wait for parallel branches, then feed downstream chart/export steps
• Approval — pause a run until an operator approves or rejects in the admin UI

Templates: built-in flows such as Patent Multi-Dimension Report (search → parallel analysis → merge → chart/Excel/Word/PPT). Save any workflow as My template and instantiate it again with one click. Skills bind by skill_name, so the same graph works across tenants after skills are installed.

Example: one patent-search skill can appear in many nodes with different payloads — command: search for retrieval, command: analysis + dimension: applicant|ipc|… for parallel analytics, then patent-report for export.

Getting started

1. make setup && make dev (or make docker-up-lite)
2. Install skills (e.g. patent-search, patent-report) under Admin → Skills
3. Open Admin → Workflows — use a template or build a graph, then Run once
4. For scheduled runs: start the worker — make dev-worker or set WORKER_ENABLED=true in .env

Beta — core paths are production-ready for validation; graph DSL and UI may evolve.
Details: Workflow orchestrator · Product tour — Workflow · API: docs/api.en.md#workflows-beta

git clone https://github.com/windinwing/mchat.git
cd mchat

make docker-up-lite
# If Docker requires sudo on your machine, the script will use it automatically.

# Admin UI:  http://localhost:5173/admin
# API docs:  http://localhost:3001/docs
# Landing:   http://localhost:5173/

Equivalent manual command (creates ops/docker/.env first):

cp ops/docker/.env.example ops/docker/.env
docker compose -f ops/docker/docker-compose.lite.yml --env-file ops/docker/.env up -d --build

Default admin credentials (created on first startup): admin / admin123
Change the password under Admin → Users after sign-in. Override via ADMIN_USERNAME / ADMIN_PASSWORD in .env. Set SHOW_BOOTSTRAP_CREDENTIALS=false to hide the hint on the login page in production.

<script
  src="http://localhost:5173/widget-loader.js"
  data-mchat-url="http://localhost:3001"
  data-agent-id="YOUR_AGENT_ID"
  data-primary-color="#3b82f6"
  data-welcome-message="Hello! How can I help you?"
  data-bot-name="Assistant"
></script>

Prerequisites (new machine):

make uses bash + source venv. Do not run make setup, make install, or make dev with sudo.
For short CLI: make install, then source scripts/env.sh, or ./bin/mchat.

After git pull (pick one path):

# Path A — local hot reload (needs Node 20+)
git pull
make setup && make dev

# Path B — Docker full stack
git pull
make docker-up-lite

First clone:

git clone https://github.com/windinwing/mchat.git
cd mchat
make setup && make dev          # or: make docker-up-lite

Docker / MySQL commands:

MySQL (single lite config):

• Credentials: mchat / mchat123, database mchat, host port 3307 (see ops/docker/.env).
• make setup syncs DATABASE_URL in src/backend/.env automatically.
• Existing MySQL: MCHAT_SETUP_MYSQL=0 make setup and set DATABASE_URL yourself.
• Access denied for user 'mchat': make db-docker-reset-lite && make setup.
• make dev stops Docker frontend/backend if they occupy port 5173 (MySQL is kept).

Step by step (without full make setup):

make install
make db-mysql-dev
make dev

make test
make lint
source scripts/env.sh && mchat run
# or: ./bin/mchat run

mchat/
├── src/
│   ├── backend/          # FastAPI (Python 3.12+)
│   │   └── app/
│   │       ├── api/      # REST routes
│   │       ├── bot/      # Bot engine + LLM providers
│   │       ├── knowledge/# RAG + Milvus
│   │       ├── skill/    # Skill system
│   │       ├── worker/   # Schedules & workflow jobs
│   │       └── channels/ # WeChat & channel adapters
│   └── frontend/         # React + Vite + Tailwind
│       └── src/
│           ├── i18n/     # zh / en (react-i18next)
│           └── pages/    # Landing + admin console
├── skills/               # Skill packages
├── channel_templates/    # Vertical channel templates (patent, medical, etc.)
├── docs/                 # Architecture, API, deployment, roadmap
├── ops/docker/           # Docker Compose
└── Makefile

See docs/api.md or /docs (Swagger) after startup.

The admin console and landing page support English and 简体中文. Language preference is stored in localStorage (mchat_lang). Switch language from the header or sidebar.

After make install:

source scripts/env.sh
mchat run
mchat skill list

Or: ./bin/mchat run. Prefer make dev for daily work.

• Supports standard frontmatter SKILL.md skill packages
• Supports OpenClaw-style SKILL.md locale blocks (auto config.i18n for UI display names)
• Admin can install skills by zip upload or URL (/api/skills/install-url)
• CLI supports direct URL or ClawHub name install, for example: mchat skill install patent-search
• Premium skill packs: Vertical channels can bundle specialized skills as value-added services
• Workflow reuse: One skill (e.g. patent-search) can appear in multiple DAG nodes with different command / dimension payloads

Backend: FastAPI, SQLAlchemy 2.0, Milvus, OpenAI / Anthropic SDKs, JWT, Loguru

Frontend: React 19, TypeScript, Vite, Tailwind CSS 4, Zustand, react-i18next, React Flow

MIT License

Contributions are welcome. See CONTRIBUTING.md if present.