OpenMAIC: The Open-Source Multi-Agent Interactive Classroom

Education is the canonical use case for AI agents: a patient teacher that explains, a curious classmate that asks questions, and a rich set of interactive materials — slides, quizzes, simulations — all generated on demand. OpenMAIC (Open Multi-Agent Interactive Classroom) from Tsinghua University’s MAIC lab is the open-source platform that delivers exactly this. Describe a topic or upload a document, and OpenMAIC generates a full lesson in minutes — with AI teachers who lecture and draw on a whiteboard, AI classmates who discuss and ask questions, and interactive scenes that include 3D visualizations, games, and project-based learning.

What Is OpenMAIC?

OpenMAIC is a Next.js 16 web application (TypeScript, pnpm monorepo) that turns any topic or document into a multi-agent classroom experience. The core idea: use multi-agent orchestration to generate a structured lesson — slides, quizzes, interactive simulations, and project-based learning activities — and then deliver it through AI agents who speak, draw, and discuss in real time.

Highlights

• One-click lesson generation — Describe a topic or attach your materials; the AI builds a full lesson in minutes
• Multi-agent classroom — AI teachers and peers lecture, discuss, and interact with you in real time
• Rich scene types — Slides, quizzes, interactive HTML simulations, and project-based learning (PBL)
• Whiteboard and TTS — Agents draw diagrams, write formulas, and explain out loud
• Export anywhere — Download editable .pptx slides, interactive .html pages, or .mp4 videos
• OpenClaw integration — Generate classrooms from Feishu, Slack, Telegram, and 20+ messaging apps

Lesson Generation Pipeline

The pipeline from topic to classroom has five stages, each with LLM-driven generation and optional per-stage model routing.

1. Document Parsing — Multi-format upload (PDF, DOCX, PPTX, audio, video, image) with audio/video extraction, AliDocMind, MinerU, and a lexical retrieval foundation (RAG).
2. Outline Generation — An LLM generates the lesson outline, which is editable before full generation. Language is auto-inferred from the input.
3. Per-Stage Model Routing — Optionally use different LLMs for the outline vs. scene generation stages via DEFAULT_MODEL config.
4. Scene Assembly — The DSL renderer generates four scene types: slides, quizzes, interactive HTML simulations, and PBL activities.
5. MAIC Editor — Pro Mode lets you edit slides directly (drag, resize, rotate, multi-select) or use “Edit with AI” with validated JSON Patch edits and multi-session history.
6. Classroom Delivery — AI teacher and classmate agents deliver the lesson with TTS, whiteboard, real-time discussion, and immersive mode.

Export Options

Format	Description
.pptx	Editable PowerPoint slides
.html	Interactive HTML pages (offline-ready)
.mp4	One-click video export via the render service (CPU resource profiles, bounded chunk executor)

Multi-Agent Orchestration

The multi-agent design is what makes OpenMAIC feel like a real classroom rather than a slide generator:

• AI Teacher Agent — Lectures, explains concepts, draws on the whiteboard, and uses TTS to speak aloud
• AI Classmate Agents — Participate in discussions, ask questions, and interact with both the teacher and the user in real time

Agents operate across four scene types:

Scene Type	Description
Slides	Editable outline with PPTX/HTML export
Quiz	Questions with a completion page and persistent quiz state
Interactive HTML	3D visualizations, simulations, games, mind maps, online programming
Project-Based Learning (PBL)	v2 classroom UI with a vocational-learning task engine

Provider Matrix: 15+ LLM Providers + Local AI

OpenMAIC supports an extensive provider matrix, covering both cloud and local AI.

Cloud LLM Providers

Provider	Models
OpenAI	GPT-5.5, GPT-5.6
Anthropic	Claude Opus 4.8, Claude Sonnet 5
Amazon Bedrock	Claude (managed, AWS credentials)
Google Gemini	Gemini 3 Flash (recommended), Gemini 3.1 Pro
DeepSeek	DeepSeek-V4
Qwen	Qwen3.7 Plus/Max
Kimi	K2.7 Code
MiniMax	M2.7, M3
Grok (xAI)	Grok models
GLM (Zhipu)	GLM-5.1, GLM-5.2
OpenRouter	100+ models via single API
Tencent	Hunyuan/TokenHub
Xiaomi	MiMo (Token Plan)
Doubao	Doubao models

Local AI Providers

Provider	Capabilities
Ollama	Local LLM inference
Lemonade	Local LLM + image generation + TTS + ASR (OpenAI-compatible, no API key)
FunASR	Local speech recognition (SenseVoiceSmall, Paraformer, Fun-ASR-Nano with vLLM)
VoxCPM2	Voice cloning with auto-generated voices
ComfyUI	Image generation

Configuration

Configure providers via .env.local or server-providers.yml:

# OpenAI
OPENAI_API_KEY=sk-...
DEFAULT_MODEL=openai:gpt-5.5

# Anthropic
ANTHROPIC_API_KEY=sk-ant-...

# Google Gemini (recommended)
GOOGLE_API_KEY=...
DEFAULT_MODEL=google:gemini-3-flash-preview

# Amazon Bedrock
BEDROCK_REGION=us-east-1
BEDROCK_MODELS=us.anthropic.claude-sonnet-5,us.anthropic.claude-opus-4-8
DEFAULT_MODEL=bedrock:us.anthropic.claude-sonnet-5

# Local (Lemonade - no API key needed)
LEMONADE_BASE_URL=http://localhost:13305/v1
TTS_LEMONADE_BASE_URL=http://localhost:13305/v1
ASR_LEMONADE_BASE_URL=http://localhost:13305/v1
IMAGE_LEMONADE_BASE_URL=http://localhost:13305/v1

The recommended model is Gemini 3 Flash for the best balance of quality and speed. For highest quality (at slower speed), use Gemini 3.1 Pro.

OpenClaw Integration: Classrooms from Your Chat App

OpenMAIC integrates with OpenClaw, letting you generate classrooms directly from Feishu, Slack, Discord, Telegram, and 20+ other messaging apps — zero local setup required.

How It Works

1. Install the skill: clawhub install openmaic or ask your Claw to “install OpenMAIC skill”
2. Pick a mode:
   • Hosted mode — Get an access code at open.maic.chat, no local setup needed
   • Self-hosted — The skill walks you through clone, config, and startup step by step
3. Tell your assistant “teach me quantum physics” — done!

Quick Start

Prerequisites

• Node.js >= 20
• pnpm >= 10

Installation

git clone https://github.com/THU-MAIC/OpenMAIC.git
cd OpenMAIC
pnpm install

Configuration

cp .env.example .env.local

Fill in at least one LLM provider key, then run:

pnpm dev

Docker Deployment

OpenMAIC ships with a Dockerfile and docker-compose.yml for containerized deployment, including a Postgres stack for server-backed persistence.

Version History

Version	Date	Key Changes
v1.0	Aug 27, 2026	Brand-new Pro Workbench UI (three-pane chat + preview), course-level planning (consecutive lessons), full-component editing, high-fidelity PPTX import, audio/video-to-lesson, voice cloning (open-source), all-new Skill System (deep research, PPT style reuse, teaching-style transfer, diverse teaching styles, custom skills)
v0.3.2	Aug 14, 2026	Video export hardening, server-backed persistence, @openmaic/generation package, 4 new locales, Bedrock/Atlas Cloud/Claude search providers, FunASR ASR
v0.3.1	Jul 21, 2026	One-click MP4 video export, Postgres storage, direct slide manipulation, “Edit with AI” Pro-mode, expanded document parsing, Azure OpenAI/SearXNG/ComfyUI providers, GPT-5.6
v0.3.0	Jun 28, 2026	PBL v2, “Edit with AI” editor agent, @openmaic/* SDK family on npm, per-stage model routing, GLM-5.2/Kimi K2.7/Qwen3.7, Korean locale, AGPL-3.0 to MIT relicense
v0.2.2	Jun 2, 2026	MAIC Editor v0, editable outline, offline classroom export, Brave/Baidu/Bocha/MiniMax search, Claude Opus 4.8/MiniMax M3/Gemini 3.5 Flash, zh-TW and pt-BR locales
v0.2.1	Apr 26, 2026	VoxCPM2 TTS with voice cloning, per-model thinking config, completion page, DeepSeek-V4/GPT-5.5/GPT-Image-2/Xiaomi MiMo
v0.2.0	Apr 20, 2026	Deep Interactive Mode — 3D visualization, simulations, games, mind maps, online programming
v0.1.1	Apr 14, 2026	Automatic language inference, ACCESS_CODE auth, classroom ZIP export/import, custom TTS/ASR, Ollama support
v0.1.0	Mar 26, 2026	Discussion TTS, immersive mode, keyboard shortcuts, whiteboard enhancements

SDK Family

OpenMAIC publishes the @openmaic/* SDK family to npm:

• @openmaic/generation — Lesson generation engine
• @openmaic/dsl — Domain-specific language for scene definitions
• @openmaic/renderer — Scene renderer
• @openmaic/importer — Document import and parsing

The SDKs enable secondary development (二开) — building custom classroom experiences on top of OpenMAIC’s orchestration layer.

Key Design Decisions

Why multi-agent instead of a single LLM? A classroom is inherently multi-agent: a teacher explains, a classmate asks questions, and the interaction between them creates the learning experience. A single LLM cannot replicate this dynamic.

Why a DSL renderer for scenes? The DSL separates scene definition from rendering, enabling the same lesson to be exported as PPTX, HTML, or MP4 without regeneration. It also enables the MAIC Editor to manipulate scenes structurally rather than as flat content.

Why per-stage model routing? Different stages have different quality/speed tradeoffs. Outline generation benefits from a strong reasoning model, while scene assembly may be faster with a lighter model. Per-stage routing lets you optimize cost without sacrificing quality where it matters.

Why MIT license? OpenMAIC started as AGPL-3.0 but relicensed to MIT in v0.3.0 to encourage adoption and secondary development, particularly in enterprise environments where AGPL is often a blocker.

Further Reading

• GitHub: THU-MAIC/OpenMAIC — MIT license
• Live Demo
• Changelog
• Contributing Guide
• ComfyUI Setup
• FunASR Deployment
• VoxCPM2 — Voice cloning TTS

Summary

OpenMAIC is the most comprehensive open-source AI classroom platform available: multi-agent orchestration with AI teachers and classmates, four scene types (slides, quizzes, simulations, PBL), 15+ LLM providers including local options (Ollama, Lemonade, FunASR), export to PPTX/HTML/MP4, an editable DSL with the MAIC Editor, an npm-published SDK family for secondary development, and OpenClaw integration for generating classrooms from any messaging app. Built on Next.js 16 with TypeScript and MIT-licensed, it is both a production-ready teaching tool and a platform for building custom interactive learning experiences.

Enjoyed this post? Never miss out on future posts by following us