AI & DevelopmentOpen SourceDeveloper Tools

OpenMAIC: Tsinghua’s Multi-Agent AI Classroom Is Now Open Source

Tsinghua University’s THU-MAIC lab just open-sourced a classroom that runs itself. OpenMAIC is not an AI tutor you ask questions — it is a live, generative classroom where a teacher agent lectures on a virtual whiteboard, classmate agents argue and push back in real time, and a TA agent fills the gaps. Upload a topic or a document, and the system builds a full lesson in roughly 30 minutes. As of today it has 25,000+ GitHub stars and is sitting at the top of trending.

Not a Chatbot With Slides

The difference between OpenMAIC and asking Claude to explain something is structural. When you use a standard AI tutor, you are responsible for the design: what to ask first, when to drill deeper, when to test yourself. That meta-work is exhausting, and most people skip it.

OpenMAIC inverts this. A Plan Agent structures the pedagogical flow before the session even starts. Then a Teacher Agent delivers it — with voice narration, whiteboard animations, and laser pointer effects. Meanwhile, AI classmate agents ask the questions you had but did not raise, challenge assumptions, and react visibly to new ideas. A TA agent catches anything the teacher glosses over.

The result is what researchers call “social texture” — the kind of learning dynamic that MOOCs promised and never delivered because pre-recorded video cannot adapt. In a study of 700+ Tsinghua students with 100,000+ learning records, OpenMAIC achieved 84.1% satisfaction and outperformed traditional MOOCs on engagement and retention. The architecture follows SimClass, grounded in Flanders Interaction Analysis and Community-of-Inquiry frameworks.

Four Agents, One Classroom

The agent team is orchestrated through LangGraph using a Director-Generator-Director loop. The four roles:

  • Plan Agent — Runs silently before the session, designing lesson structure and pacing
  • Teacher Agent — Delivers the lesson with TTS narration, whiteboard drawing, and UI highlights
  • Classmate Agents — Participate live: ask questions, disagree, emotionally react
  • TA Agent — Handles follow-up questions, fills conceptual gaps, keeps the flow moving

The tech stack is Next.js 16, TypeScript, Zustand, and PostgreSQL. The platform supports OpenAI, Anthropic Claude, Gemini, and DeepSeek as model backends — your keys, your choice.

PBL Mode: Vibe Coding With Guardrails

Version 1.4.0, released November 4, introduced Project-Based Learning v3. Instead of a lecture, you build a Python or Java project inside an in-browser IDE with a terminal. The AI tutor guides you milestone by milestone, running tests at each stage. When you finish, you get a learning report and a shareable achievement image. Projects export as a portable .maic.zip.

The team describes it as “a new way into vibe coding where you build the project and understand how it works.” That framing is deliberate — it positions PBL mode as a corrective to the pattern where vibe coding produces working code without the developer understanding why. Whether it delivers on that is worth testing.

Self-Hosting in Under 10 Minutes

The setup is genuinely straightforward. Clone, install, run:

git clone https://github.com/THU-MAIC/OpenMAIC.git
cd OpenMAIC
pnpm install
pnpm db:up          # starts local PostgreSQL
# Add LLM keys to .env.local
pnpm dev

For production: docker compose up --build. One-click Vercel deployment is also available. The getting started docs are clean. The license is MIT — no commercial use restrictions.

Limitations Worth Knowing

A few things to have in mind before you run this in production. First, you bring your own LLM keys — costs scale with usage, especially if classmate agents are running multiple calls per session. Second, there is no built-in user identity system; authentication is a browser cookie or a shared access code, which is fine for personal use but needs work for multi-user deployments. Third, the release pace has been fast enough that some updates have required database backups before upgrading — keep that in mind if you depend on the hosted demo at open.maic.chat.

The Takeaway

OpenMAIC is the most interesting application of multi-agent orchestration to education that has shipped this year. The architecture — autonomous agent roles collaborating in real time to simulate a learning environment — is a template worth stealing for other domains. If you are building in ed-tech, working on agent-based UX patterns, or just want to actually learn something deeply instead of querying a chatbot, it is worth an hour of your time. The repo and the live demo are both free.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *