OpenMAIC Review: Tsinghua's Free AI Classroom, 42,832 Stars Strong

What OpenMAIC Is

OpenMAIC is an open-source platform from Tsinghua University's MAIC team that turns a topic, or a document you upload, into a full interactive classroom. You type "photosynthesis" or drop in a PDF, and the software builds a lesson of slides, quizzes and interactive simulations, then delivers it through AI teachers and AI classmates who lecture, argue, draw on a whiteboard and answer questions in real time. It is the most ambitious open-source attempt at generated teaching I have seen, and the numbers back up the attention: the repository passed 42,800 GitHub stars in its first eight months, with version 1.4.1 landing on November 6, 2026.

The project grew out of the MAIC team's research on LLM-driven teaching, published in the Journal of Computer Science and Technology in 2026 under the title "From MOOC to MAIC." The idea is a direct answer to the weakness of recorded video courses: a video cannot notice that you are lost, cannot call on you, and cannot change its explanation. OpenMAIC replaces the recording with a small cast of agents that can. The teacher lectures with voice narration, a classmate asks the question you were about to ask, and the whiteboard fills in with a diagram drawn step by step.

That scope is also what separates it from the nearest open-source rival. DeepTutor, from the University of Hong Kong, is a tutor for your own books: it reads your material and quizzes you on it. OpenMAIC builds the whole classroom around the material, with multiple agents, simulations and project work. Both are free and open source. The difference is whether you want a tutor or a school.

What a Lesson Actually Looks Like

Generation runs in two passes. First the system reads your topic or document and writes a structured outline. Then each outline item becomes a scene: a slide deck with narration, a quiz, an interactive HTML simulation, or a project-based learning activity. The whole lesson is generated on the server, so closing the tab or restarting the service no longer kills a run in progress, a fix that arrived with version 1.2.0.

The classroom itself has four parts that matter. Slides come with voice narration, spotlight effects and laser-pointer animation, close to a real lecture. Quizzes cover single choice, multiple choice and short answers, with grading and feedback from the AI in real time. Interactive simulations are HTML experiments the learner can manipulate, from physics demos to flowcharts. Project-based learning, rebuilt as PBL v3 in version 1.4.0, gives the learner a role, a structured project with milestones and a personal learning report at the end.

Discussion is where the multi-agent design shows. Agents start discussions on their own, the learner can jump in at any point or get called on, and a roundtable mode has several agents with different personas debate a topic while one of them sketches on the whiteboard. A Q&A mode answers free questions with slides, diagrams or whiteboard drawings. When the lesson is done, it exports as a PowerPoint file that keeps every scene, a single offline HTML file that plays the whole lesson, or a classroom ZIP to move it to another deployment. MP4 video export exists too, through a separate render container running Chromium and FFmpeg.

What Is New in Version 1.4.1

Version 1.4.1 shipped on November 6, 2026, two days after 1.4.0, and it is a stability release rather than a feature drop. Background syncs no longer discard unsaved edits, empty a project or delete a page an agent just added, and courses open faster on slow connections. The headline feature is one release back: version 1.4.0, released November 4, added project-based learning v3, where the learner builds a real project with an AI tutor guiding requirements, design and implementation, and gets staged feedback at each milestone.

The release before that, 1.3.0 from October 29, added a knowledge base that keeps every upload until deleted, organized in folders and parsed once for every conversation, plus a Pro workbench upgrade. The Pro workbench is a conversational course-building surface with 24 built-in skills covering curriculum planning, research, lecture and workshop styles, and it is off by default. The cadence matters here: three releases in nine days, each with breaking-change notes, tells you this is a fast-moving project, and that you should read the release notes before upgrading. Version 1.3.0 upgrades are one-way, with no rollback below 1.3.0.

How to Run It

There are two ways in. The fastest is the hosted demo at open.maic.chat, which runs the live build and needs an access code from the site. No install, no keys, and it is the right way to decide whether the classroom format suits you before you commit to hosting it.

Self-hosting needs more than a laptop and a free evening. The requirements are Node.js 22.19 or newer, pnpm 10 or newer, and PostgreSQL 16, and generation runs as a long-lived server process with the database behind it. The Docker route is the shorter path: copy the example environment file, add your API keys, and run docker compose up --build, which starts two containers, the app and PostgreSQL, and serves the app on localhost port 3000. The manual route is git clone, pnpm install, pnpm db:up for a development database, then pnpm dev.

Configuration lives in a single file, openmaic.yml, where you declare providers and assign a model to each job: one model writes the outline, another writes slide content, another handles speech, web search and video. The supported list is long: OpenAI, Azure OpenAI, Anthropic, Amazon Bedrock, Google Gemini, DeepSeek, Qwen, Kimi, MiniMax, Grok, OpenRouter, TokenDance, Doubao, Tencent Hunyuan, Xiaomi MiMo and GLM, plus Ollama and any OpenAI-compatible API. Local options exist too: Lemonade runs local models with no API key, FunASR transcribes locally, and VoxCPM2 provides self-hosted text-to-speech with voice cloning.

One setting deserves attention before anything else. The Docker Compose file is a personal installation: it runs in single-user mode and listens on localhost only. The moment you publish it to a network, set a long random ACCESS_CODE first. Without one, the project's own documentation is blunt: anyone who can reach the port owns, edits and can delete the whole library. There is no account system unless you build one through the project's auth hooks.

What It Costs

The software costs nothing. OpenMAIC is licensed under the MIT License, confirmed in the repository's license file, which permits commercial use free of charge. There is no paid edition, no subscription tier and no per-seat price. The Pro workbench is a feature inside the same free codebase, not a plan.

Your bill is the AI underneath. Every lesson consumes tokens from the providers you connect, across text, images, narration, video and web search, and the total depends entirely on which models you pick. At launch, coverage of the project reported a full course generated in about 30 minutes for less than $2 in model usage, roughly £1.49 or C$2.81 at November 8, 2026 market rates. Treat that as a launch-era figure, not a promise: it dates from March 2026, and premium models cost far more than the budget ones. The project's own documentation recommends a fast long-context model such as DeepSeek V4.1 Flash as the default, and independent coverage of the project warns that long sessions on premium models get expensive.

That trade-off is the real pricing story. You can keep a lesson cheap by pointing every slot at budget models, or you can spend more for better writing, narration and images. There is no meter inside OpenMAIC that warns you first. Set spending limits at your provider, not in the app, and check the first few lessons against your provider's billing page before you hand the classroom to students.

OpenMAIC vs DeepTutor

The two most-starred open-source AI teaching projects of 2026 come from two universities and take two different routes to the same goal. DeepTutor, from the University of Hong Kong, is a tutor: give it your books and it teaches from them. OpenMAIC, from Tsinghua, is a classroom: give it a topic and it builds the teacher, the classmates, the whiteboard and the coursework.

Pick DeepTutor if you already have the material and want a tutor that works through it. Pick OpenMAIC if you want the material turned into a taught course, with other agents in the room and something to export at the end. If you work from a chat app rather than a server, OpenMAIC has one more route: its skill package works with OpenClaw, the assistant you text from WhatsApp, and with Codex, DeepSeek and WorkBuddy, so you can ask for a classroom from Feishu, Slack, Telegram or WhatsApp and get the link back when it is ready.

The Catches

OpenMAIC is strong software with rough edges that its own documentation does not hide. The security posture is the first. There is no authentication unless you set ACCESS_CODE, and the code is a shared site password, not per-user accounts. The Compose deployment listens on localhost for a reason. Do not expose it to a network without the access code and a random database password.

The second is the upgrade treadmill. Serverless hosting stopped at version 1.1.x; from 1.2.0 the app needs a long-running server and PostgreSQL, and Vercel deployments are no longer supported. Upgrades carry breaking changes and one-way migrations, and the project has shipped a string of security fixes, from path validation and stored HTML to server-side request forgery patches across versions 1.0.1 through 1.1.2. Keep a self-hosted instance current, and read the changelog before each upgrade.

The third is accuracy. Nothing in OpenMAIC checks whether a lesson's facts are right. The agents generate confidently, and a wrong explanation delivered by a charming AI teacher is still wrong. The Tsinghua trials cited in launch coverage involved more than 500 students and over 100,000 interaction records, with 84.1% satisfaction reported, but those are the team's own figures, not an independent evaluation, and satisfaction is not the same as correctness. Review every lesson before students see it, especially in subjects where a wrong fact has consequences.

Two smaller points from independent coverage are worth knowing. One bundled package carries an LGPL-3.0 license, so check the terms before redistributing the software commercially. And the classroom chat needs a model with tool calling, while the agent slot cannot set a thinking-effort level, so some model features stay out of reach.

Who Should Use It

OpenMAIC fits teachers and trainers who want a course built from their own material and are comfortable running a small server, or at least a Docker container. It fits developers building learning products on top of the project's SDKs, and it fits curious learners who want to see what a multi-agent classroom feels like before committing to a paid platform. The hosted demo is the right first step for all three.

It does not fit anyone who wants a finished product with accounts, billing and support. There is no hosted pricing, no service-level promise, and no team to call when a lesson generation fails. If you need a classroom that simply works for a hundred students tomorrow, a managed platform is the honest choice. If you want the most complete open-source attempt at AI teaching, and you accept the setup and the review work, OpenMAIC is the one to try first.

Is OpenMAIC really free?

Yes, the software is free under the MIT License, with no paid edition or subscription. You pay only for the AI models you connect, plus whatever hosting you use. The hosted demo at open.maic.chat is free to try with an access code, and no hosted price is published.

Do I need a powerful computer to run OpenMAIC?

Not for the app itself. It needs Node.js 22.19 or newer, pnpm 10 or newer, and PostgreSQL 16, and Docker handles all three. The heavy work happens at your model provider. Only local models through Ollama or Lemonade need your own GPU for reasonable speed.

Can OpenMAIC teach in English?

Yes. The interface, documentation and user guide are available in English as well as Chinese, and the lessons are generated in the language you ask for. The project comes from a Chinese university team, but English is a first-class option.

Is OpenMAIC safe to put on the internet?

Not by default. Without an access code, anyone who can reach the server can read, edit and delete the whole course library. Set a long random access code and a random database password before publishing it beyond localhost, and add real authentication before serving other people.

How is OpenMAIC different from DeepTutor?

DeepTutor is a tutor that teaches from your own books. OpenMAIC builds a full classroom: AI teachers and classmates, a whiteboard, quizzes, interactive simulations and project work, with exports to PowerPoint and offline HTML. Both are free and open source, and the choice is tutor versus classroom.

Can I use OpenMAIC without coding?

Mostly. The hosted demo needs no setup at all. Self-hosting needs a terminal for the Docker commands and the configuration file, but no programming. The OpenClaw skill goes further, letting you request a classroom from a chat app and receive the link when it is ready.

Sources

• OpenMAIC on GitHub (THU-MAIC/OpenMAIC) — README, releases, license and configuration, read November 8, 2026
• OpenMAIC v1.4.1 release notes — November 6, 2026
• OpenMAIC v1.4.0 release notes (PBL v3) — November 4, 2026
• OpenMAIC live demo — verified loading November 8, 2026
• "From MOOC to MAIC," Journal of Computer Science and Technology (2026)
• ReviewRamble: OpenMAIC GitHub repo review — answers checked October 31, 2026
• Frontier Now: OpenMAIC AI classroom — October 18, 2026
• AIbase: Tsinghua open-sources OpenMAIC — launch coverage, March 2026
• AllClaw: OpenMAIC review — March 18, 2026
• RateMorph: USD to GBP mid-market rate — November 8, 2026