OpenMAIC: An open-source multi-agent interactive classroom by the Tsinghua MAIC team. Provide a topic or upload materials to generate slides, quizzes, interactive HTML simulations, and project-based learning activities. Lessons are taught and discussed with you by AI teachers and AI classmates who can speak, draw diagrams, and write formulas on a whiteboard.
Stars 42191 (+116 / 24h)
Reasons
- v1.3.0 released on 10/29 added a knowledge base: uploaded materials remain in each account's knowledge base, organized by folder, parsed only once, and referenced via @ in the chat box; agents can read, search, and organize them. The Pro workspace provides a work preview before generating the first page, and classroom pages allow switching between learning and editing modes.
- Starting with v1.2 in mid-10, generation moved to the server side, so closing tabs or restarting does not interrupt it; classrooms can be exported as PPTX files preserving each scene, HTML files for offline playback of the entire lesson, or classroom ZIPs for migration.
- Includes OpenMAIC skills, allowing classroom generation within agent workspaces like OpenClaw, Codex, DeepSeek, and WorkBuddy, and initiation from Feishu, Slack, and Telegram; the corresponding paper was published in JCST'26
Usage
- Requires Node.js 22.19+, pnpm 10+, and PostgreSQL 16 (for local development, use
pnpm db:upto start one in Docker); aftergit clone, runpnpm install - Copy
.env.exampleto.env.localandopenmaic.example.ymltoopenmaic.yml; at minimum, fill inOPENAI_API_KEY. Inopenmaic.yml, specify models for "slots" such as outline, content, voiceover, and search; supports OpenAI, Anthropic, DeepSeek, Qwen, Kimi, MiniMax, GLM, Ollama, and any OpenAI-compatible interface - If you don't want to deploy, you can first use the official online demo at open.maic.chat
Risks
- The authors state that enabling images, voiceover, video, and search provides the most complete experience but incurs corresponding model costs; classroom chat requires models that support tool calling.
- Before upgrading to v1.3.0, all old instances must be stopped, and rollback to versions below 1.3.0 is not possible; starting with v1.2, a long-running server is required, and Vercel deployments can only stay on 1.1.x; license is MIT