OpenMAIC
Tsinghua University (THU-MAIC) · OpenMAIC: Open-Source Multi-Agent Classroom That Turns Topics and Documents into Interactive Lessons
OpenMAIC is a free, MIT-licensed platform from Tsinghua University's THU-MAIC group that turns a topic or uploaded material into an interactive classroom: slides, quizzes, simulations, and project-based activities, taught by AI teachers and AI classmates who narrate, draw on a whiteboard, and discuss with you. It suits self-learners, educators, and course builders who want generated lessons they can host, edit, and export.
Best for
Self-learners, teachers, and course or training builders who want AI-generated, multi-agent interactive lessons from a topic or their own materials, and technical teams that want to host, modify, or embed that capability under an MIT license
Not ideal for
Schools that need a managed, supported service with user accounts, LMS integration, and published terms, people who only want quick cited answers or summaries from documents, and anyone unwilling to run PostgreSQL and manage model API keys
Who it's for
Self-learners, educators, instructional designers, and developers who want to generate, host, or build on multi-agent interactive lessons
Capabilities
- Generates a full lesson from a one-line topic or uploaded material through an outline stage and a scene stage; PDF and text work out of the box, while Word, PowerPoint, spreadsheet, image, audio, and video inputs depend on the document and media extractors you configure
- Scene types: narrated slides with spotlight and laser-pointer effects, quizzes (single choice, multiple choice, short answer) with AI grading, HTML simulations, and project-based learning with roles, milestones, and deliverables
- Deep Interactive Mode with 3D visualizations, simulations, knowledge games, mind maps, and in-browser programming, where the AI teacher can operate the interface to guide the learner
- Multi-agent classroom: AI teachers and classmates lecture, start discussions, call on you, hold roundtable debates with different personas, and draw equations and diagrams on a shared whiteboard
- Classroom chat on an agent loop (v1.1.0): point at a slide element, interactive component, or whiteboard drawing and ask about it, and the teacher can read the lesson, sample an experiment's live state, and search the web
- Pro agent workbench (v1.0.0): a chat agent that plans multi-lesson curricula, creates and reorders pages, edits narration, imports .pptx decks with layout preserved, and generates images, video, and narration audio
- Durable, PostgreSQL-backed agent sessions that survive restarts and can be cancelled, resumed, and steered mid-run, plus 24 built-in teaching and slide-craft skills and per-owner custom skills
- Exports to editable PowerPoint (with charts and LaTeX formulas), self-contained interactive HTML, a classroom ZIP that plays offline on an intranet instance, and MP4 through an optional render service
- Provider-neutral model configuration in openmaic.yml with a model per generation step, covering OpenAI, Azure OpenAI, Anthropic, Amazon Bedrock, Google Gemini, DeepSeek, Qwen, Kimi, MiniMax, Grok, OpenRouter, GLM, Doubao, and any OpenAI-compatible API
- Local options: Ollama and Lemonade for LLMs, Lemonade for image, TTS, and speech recognition, FunASR for speech recognition, and self-hosted VoxCPM2 text-to-speech with voice cloning
- OpenMAIC skill (SKILL.md) for OpenClaw, Codex, DeepSeek, WorkBuddy, and other agent workbenches, so classrooms can be generated from Feishu, Slack, Telegram, or an IDE
- HTTP API for asynchronous classroom generation from a requirement and up to five uploaded materials
- Interface in 12 locales across 11 languages, including English, Simplified and Traditional Chinese, Japanese, Korean, Spanish, French, German, and Arabic
Limitations
- Self-hosting requires PostgreSQL 16 (the server refuses to start without DATABASE_URL), Node.js 22.19+ and pnpm 10+, or the bundled Docker Compose stack
- You bring your own model provider keys or local models, and the project recommends enabling image, voice, video, and web search providers for the best lessons, which adds setup and per-use cost
- No built-in user accounts: identity is an anonymous browser cookie, a single-user mode, or one shared team owner behind an access code, and real multi-user logins require writing your own authentication method
- ACCESS_CODE is unset by default and the app is then fail-open, so a deployment reachable by others must be protected before it is exposed
- Course reads are capability-by-id: anyone who can reach the server and knows a course id can read that course, though only its owner can change it
- The Pro agent workbench is off by default and must be enabled with build and runtime flags plus a model that supports tool calling
- On the hosted Live Demo at open.maic.chat, access codes require signing in and generation through the skill or API is capped at 10 per day; the web interface's quota and the demo's terms are not documented in the repository
- September 2026 brought five security releases (v1.0.1 to v1.0.3, v1.1.1, v1.1.2), several with breaking or behavior changes, so self-hosters need to follow upgrade notes closely
- MP4 export needs a separate render-service container with Chromium and FFmpeg; without it, export falls back to a project ZIP for local rendering
- No integration with learning management systems such as Moodle, Canvas, or Google Classroom is documented
Use cases
- Turning a research paper or report PDF into a narrated lesson with a quiz and a class discussion
- Learning a new skill from scratch, such as introductory Python, with in-browser coding exercises
- Building a multi-lesson course with the Pro workbench from existing slides, documents, and lecture recordings
- Creating interactive simulations or 3D visualizations to explain a physics or engineering concept
- Generating a classroom from a chat message in Feishu, Slack, or Telegram through the OpenClaw skill
- Exporting a generated lesson to PowerPoint or an offline classroom package for an intranet or air-gapped school network
Our take
OpenMAIC is an unusually complete open-source take on an AI-run classroom: it does not just answer questions about a document, it turns the material into a lesson with narration, a whiteboard, quizzes, simulations, and AI classmates, and since v1.1.0 the teacher can reason about what the learner is looking at. That makes it a different decision from Gemini Notebook. Choose Gemini Notebook to interrogate and summarize sources with citations; choose OpenMAIC when you want to be taught, or want to produce courses others will take. The catch is operational. Running it well means PostgreSQL, a stack of model and media keys, and care with access codes and identity, because it ships without user accounts and is fail-open until you set a code. Try the Live Demo first, then self-host only if you are ready to maintain a fast-moving project.
Who should use it
Self-learners who want an interactive lesson instead of a summary, teachers and trainers who want a first draft of a course from their own slides and documents, instructional designers prototyping interactive content, and developers or edtech teams who want an MIT-licensed base to host or embed.
Who should skip it
Schools and companies that need a supported product with user accounts, LMS integration, and published terms, people who just want cited answers from their documents, and anyone who does not want to operate PostgreSQL or manage provider API keys.
Strengths
- Produces a taught, interactive class with multiple agents rather than static notes or summaries
- Wide range of scene types: narrated slides, graded quizzes, simulations, games, 3D visualizations, and project-based learning
- MIT licensed with npm SDK packages, so it can be hosted, modified, and embedded commercially
- Provider-neutral, with a model per generation step and local options for LLM, voice, and speech recognition
- Agent workbench builds and revises whole courses from your own materials and slides
- Exports to editable PowerPoint, interactive HTML, offline classroom packages, and MP4
Weaknesses
- Self-hosting needs PostgreSQL, Node.js, and several provider keys for the full experience
- No built-in user accounts or LMS integration
- Insecure if exposed without setting an access code, and courses are readable by anyone with their id
- The hosted Live Demo is capped, and its pricing and terms are not documented in the repository
- Rapid release pace with frequent security and breaking changes
OpenMAIC pricing
Open Source
Free
- MIT license, commercial use permitted
- Self-host with Docker Compose, Vercel plus an external PostgreSQL database, or from source
- Bring your own model, media, and search provider keys, or run local models
Free tier limits: The software has no paid tier. The hosted Live Demo at open.maic.chat limits generation through the skill or API to 10 per day (with an access code you get by signing in); its web interface has a separate quota that is not documented in the repository.
Note: Self-hosted costs come from the model, image, video, voice, and search providers you configure, or only your hardware if you use local providers such as Ollama, Lemonade, and FunASR. The repository documents no pricing or paid plan for the hosted Live Demo. The bundled mathml2omml package is LGPL-3.0-or-later rather than MIT.
Technical specs
Modalities
Text, Documents (PDF, Word, PowerPoint, spreadsheets), Image, Audio, Video, Voice (TTS and speech recognition)
API pricing
Free self-hosted HTTP API; model, media, and search usage is billed by the providers you configure
Available models
Where OpenMAIC excels
Learning a dense paper as a class instead of reading it alone
Upload the PDF and OpenMAIC outlines it into scenes, has an AI teacher walk through it with narration and whiteboard drawings, and lets AI classmates raise questions, with a quiz to check understanding.
A first draft of a training course from existing materials
The Pro workbench agent plans a multi-lesson curriculum from uploaded documents, slides, and recordings, imports existing .pptx decks, and revises pages on request, and the result exports to editable PowerPoint.
Hands-on STEM explanations
Deep Interactive Mode generates simulations, 3D visualizations, and coding exercises, and since v1.1.0 the teacher can read the live state of an experiment when the learner asks about it.
Classrooms for offline or intranet networks
Exported classroom packages inline the scripts, fonts, and images that interactive scenes use, so a course can play on an air-gapped OpenMAIC instance without reaching a public CDN.
OpenMAIC vs. competitors
OpenMAIC vs. Gemini Notebook (formerly NotebookLM)
Gemini Notebook (formerly NotebookLM) is Google's hosted research assistant that answers only from the sources you add, with citations, and generates study aids such as audio and video overviews, flashcards, quizzes, mind maps, and slide decks. OpenMAIC also starts from your material or a topic, but its output is a taught lesson: AI teachers and classmates narrate slides, run quizzes and simulations, draw on a whiteboard, and discuss with you, and you can host, modify, and export it. Pick Gemini Notebook for cited research and quick study aids with no setup, and OpenMAIC for an interactive, multi-agent class or for building courses on infrastructure you control.
Frequently asked questions
What is OpenMAIC?
OpenMAIC (Open Multi-Agent Interactive Classroom) is an open-source platform from Tsinghua University's THU-MAIC group that turns a topic or uploaded documents into an interactive lesson. Multiple AI agents act as teachers and classmates, delivering narrated slides, quizzes, simulations, and project-based activities, drawing on a whiteboard, and discussing the material with you.
Is OpenMAIC free and open source?
Yes. OpenMAIC is released under the MIT license, which permits commercial use, after moving from AGPL-3.0 in v0.3.0. The software is free; if you self-host, you pay only for the model, media, and search providers you connect, or nothing beyond your hardware with local providers.
Can I try OpenMAIC without installing it?
Yes. The OpenMAIC team hosts a Live Demo at open.maic.chat. After signing in, you can generate an access code in your account settings to use it from the OpenMAIC skill or API, where generation is limited to 10 per day. The repository does not document pricing or terms for the demo.
What do I need to self-host OpenMAIC?
Node.js 22.19 or newer, pnpm 10 or newer, a PostgreSQL 16 database, and at least one LLM provider key, or the bundled Docker Compose stack, which runs the app and PostgreSQL on 127.0.0.1 by default. Set a long random ACCESS_CODE before exposing it to other machines, and enable the Pro agent workbench with its flags if you want it.
Can OpenMAIC run with local models?
Yes. It supports Ollama and Lemonade for local LLMs, Lemonade for local image generation, text-to-speech, and speech recognition, FunASR for local speech recognition, and a self-hosted VoxCPM2 text-to-speech backend with voice cloning. Lesson quality then depends on the local models you choose.
Does OpenMAIC work with OpenClaw?
Yes. OpenMAIC ships a SKILL.md skill that installs in OpenClaw with clawhub install openmaic and can be imported into Codex, DeepSeek, WorkBuddy, and other agent workbenches. It walks you through using the Live Demo or a local install, submits generation jobs, and sends you the classroom link, asking for confirmation before state-changing steps.
OpenMAIC vs Gemini Notebook: what is the difference?
Gemini Notebook (formerly NotebookLM) answers questions about your sources with citations and turns them into study aids such as audio and video overviews, flashcards, quizzes, and mind maps. OpenMAIC turns a topic or your materials into a full interactive lesson delivered by multiple AI agents, with simulations, project-based activities, and a whiteboard, and it is open source and self-hostable.
Integrations & fit
About OpenMAIC
OpenMAIC (Open Multi-Agent Interactive Classroom) is a Next.js web application from Tsinghua University's THU-MAIC group, whose research appears in the group's JCST 2026 paper "From MOOC to MAIC". You describe what you want to learn, or attach PDFs, Word, PowerPoint, spreadsheet, image, audio, or video files, and a two-stage pipeline writes a lesson outline and then turns each item into a scene: narrated slides, quizzes with AI grading, HTML simulations, or a project-based learning (PBL) activity where you take a role alongside AI agents. Deep Interactive Mode adds 3D visualizations, simulations, games, mind maps, and in-browser programming. During playback, a cast of AI agents lectures with voice, uses a spotlight and laser pointer, draws on a shared whiteboard, starts discussions, and runs roundtable debates; since v1.1.0 the classroom chat runs on an agent loop that can read the lesson, check the live state of an experiment, and search the web before answering. Version 1.0.0 added a Pro agent workbench for course builders: a chat agent that plans multi-lesson curricula, builds and revises pages from your materials, imports .pptx decks, and generates images, video, and narration, with durable server-backed sessions you can cancel, resume, and steer, and a library of built-in teaching skills (24 in the current README). Lessons export to editable PowerPoint, interactive HTML, a classroom ZIP, or MP4. A SKILL.md package lets OpenClaw, Codex, and other agent workbenches generate classrooms from chat. What sets it apart from document tools such as Gemini Notebook is that it produces a taught, multi-agent class rather than answers, summaries, and study aids, and that you can host and modify all of it. The tradeoffs: self-hosting means running PostgreSQL and bringing your own model keys (or local models), the best results need image, voice, video, and search providers configured, there are no built-in user accounts, and the hosted Live Demo is capped, with no pricing or terms documented in the repository.
Updates from OpenMAIC
Server-side requests to caller-supplied provider URLs now connect only to validated addresses and refuse redirects, and error responses no longer echo provider responses. It followed v1.1.1, which hardened MinerU Cloud document parsing.
Classroom chat moved to an agent loop by default, so learners can ask about a specific slide element, interactive component, or whiteboard drawing, and the teacher can read the lesson, check an experiment's live state, and search the web. Settings were rebuilt with a model choice per generation step and Token Plan connections.
OpenMAIC 1.0 added a Pro workbench where a chat agent plans curricula and builds and revises pages from uploaded materials, with durable server-backed sessions, 20 built-in skills at launch, provider-neutral server configuration, and pluggable storage.
Version 0.3.0 moved the project from AGPL-3.0 to the MIT license and added Project-Based Learning v2, an Edit with AI editor agent, and the @openmaic/* SDK packages on npm.
Are you the founder? Claim this listing →