OpenMAIC
Open Multi-Agent Interactive Classroom. Get an immersive, multi-agent learning experience in just one click
The Lens
By Erik Loyd, SaaS CEO and former COO/CFO of an AWS Premier Partner.
Updated Oct 2026
OpenMAIC turns a topic, or your own documents, into an interactive class: slides, quizzes, simulations and project work, taught by AI teachers and classmates who talk, draw on a whiteboard and discuss with you. It comes from Tsinghua University's MAIC lab, is backed by a peer-reviewed paper, and is MIT licensed and free.
Self-hosting got heavier in October. Version 1.2.0 moved generation to the server, so you need PostgreSQL and a long-running Node process (docker compose up gives you both); serverless hosts like Vercel stop at 1.1.x. You bring the model keys, and lessons export as PPTX or offline HTML.
There's no paid tier. Solo learners and teachers: free, you pay only for tokens. Schools and training teams: free, and MIT allows commercial embedding. Budget tokens seriously: the project's own figures put a full agent-workbench course near 3.6 million.
The catch is security. From September 6 to October 11 the project published 16 advisories, two critical: unauthenticated SSRF into cloud metadata, arbitrary file writes, stored XSS. All patched, but with ACCESS_CODE unset every route is open to whoever reaches the port. Stay current and keep it off the open internet.
Free vs Self-Hosted vs Paid
fully freeFree: Everything. MIT licensed since v0.3.0 (June 28, 2026), relicensed from AGPL-3.0. One bundled workspace package, packages/mathml2omml, stays LGPL-3.0-or-later. The THU-MAIC team also runs a hosted instance at open.maic.chat (sign-in, and access codes for the OpenClaw skill's hosted mode); it publishes no pricing.
Self-hosted: Since v1.2.0 (October 16, 2026) it needs PostgreSQL 16 and a long-running Node.js 22 server; Docker Compose ships both and binds to 127.0.0.1 in single-user mode by default. Serverless hosts such as Vercel are supported only up to 1.1.x. MP4 export needs an opt-in render-service container (Chromium plus FFmpeg). Set a long random ACCESS_CODE before exposing it to anyone else.
Paid: No paid tier and no vendor to buy from. The bill is model usage from whichever providers you configure (OpenAI, Anthropic, DeepSeek, Gemini, Bedrock, MiniMax token plans and more), or nothing beyond hardware if you run local models through Ollama or Lemonade. Usage scales hard: the project's own figures, reported by ifanr, put a classic classroom at about 358K tokens and a full agent-workbench course at about 3.6 million tokens across roughly 74 model calls. Classroom chat now requires a model with tool calling.
Free and MIT licensed with no paid tier. Your real cost is model tokens, and a full agent-workbench course runs about 3.6 million of them by the project's own figures.
What to do by team size
- Solo
- free; docker compose runs it on localhost, and you pay only for tokens
- Small team
- free; set a long ACCESS_CODE before sharing it and budget tokens per course
- Medium team
- free; run 1.2.x on Postgres with an identity mode and model allow-lists set in openmaic.yml
- Large team
- free and MIT, so embedding is fine, but you own patching, auth and a fast-moving security advisory feed
Get tools like this every Wednesday
One featured tool, three on the radar. No fluff.
Similar Tools
A low score is not a verdict on quality. Young and niche tools start low by design. How we calculate scores
Trust Signals
License: MIT License
Bundled packages/mathml2omml is LGPL-3.0-or-later.
Commercial use: ✓ Yes
About
- Owner
- MAIC (Organization)
- Stars
- 42,775
- Forks
- 6,582
- discussed
Explore Further
More tools in the directory
Featured in The Open Source Drop #24