• Demo (inline playback):
• Bilibili full demo: https://www.bilibili.com/video/BV1X163BQE5c/?share_source=copy_web&vd_source=e0705640ea2f51669a392fb07684e286

• Create sub‑agents dynamically
• Message any agent directly
• WeChat‑style chat UI: intervene at any depth anytime
• Streaming graph that visualizes collaboration in real time

• Graph shows swarm topology and live communication links
• Tree‑structured multi‑level dialogs: talk to any agent like a chat app (even deep ones)
• LLM history panel: shows the agent’s context; no more black box
• Streaming tool‑call parameters in real time

• Minimal primitives: only a few communication primitives are required (core is create + send)
• Liquid topology: topology evolves during runtime; agents “hire” sub‑agents on demand
• Flat collaboration: humans can step into any layer at any time

No complex nodes/edges abstraction — just think of “many people”:

Everyone can spawn children and talk to anyone.

With only these two capabilities, any structure can be built.

cd agent-wechat
cd backend

cp .env.example .env.local
# Fill your keys and model names in .env.local

docker compose up -d
curl -X POST http://127.0.0.1:3017/api/admin/init-db
bun install
bun dev

Open http://localhost:3017

Click init-db, then create a workspace to start.

Try: “Create 3 children, ask them to message each other, then have each create 3 grandchildren.”

https://zhuanlan.zhihu.com/p/2000736341479138182

The backend reads backend/.env.local:

• GLM_API_KEY / GLM_MODEL (I use GLM CodingPlan)
• OPENROUTER_API_KEY / OPENROUTER_MODEL (I use OpenRouter Kimi 2.5)
• See backend/.env.example for all fields