OpenMAIC v1.0.0: Open-Source AI Classroom Gets an Agent Workbench

One of the more ambitious open-source AI education projects just hit 1.0. On August 27, 2026, OpenMAIC — the Open Multi-Agent Interactive Classroom from Tsinghua University's MAIC group — shipped v1.0.0, adding a chat-first agent workbench to a project that already had roughly 26,000 GitHub stars, 4,700 forks, and 83 contributors.

OpenMAIC's pitch is not "generate a slide deck." It turns a topic or a document into a full interactive classroom: an AI teacher lectures with voice and a laser pointer, AI classmates ask questions and get called on, a shared whiteboard fills with equations and diagrams, and learners poke at 3D visualizations, simulations, and in-browser code instead of just watching. v1.0.0 makes the course-building itself agentic.

TL;DR — what people are asking

The classic pipeline vs the new workbench

OpenMAIC now has two ways to build a course.

Classic (one-click). A two-stage pipeline: you describe a topic or attach materials, an outline is generated, then each outline item becomes a rich scene — a slide, a quiz, an interactive simulation, or a project-based-learning activity. It is fast and hands-off, and it is still the default.

Pro (agent workbench). New in v1.0.0. A chat-first workspace where a course-building agent works through explicit, validated tools rather than editing an opaque blob. It can:

• Plan multi-lesson curricula, create courses and folders, rename and reorder.
• Read and search the underlying stage DSL, then atomically patch one scene at a time.
• Generate, duplicate, insert, delete, and reorder pages; edit narration and deck structure.
• Pull from session materials — uploaded documents, audio, and video, or trusted web URLs — and search the extracted text.
• Generate images, video, and narration audio through your configured providers.
• Import .pptx slides with layout preserved.

The consequential part is that workbench sessions are database-backed with leases, heartbeats, and crash resume. A run survives a worker restart, accepts follow-up instructions while it is working, and streams a replayable event history to the chat. That is the difference between a demo and something you can leave running on a real curriculum.

Twenty built-in skills

The workbench ships with 20 reusable skills covering curriculum planning, deep research, and distinct teaching styles — interactive, lecture, workshop, vocational — plus slide and stage craft, PPTX import, editing, and style reuse. Users can author their own skills, stored per owner and edited through the same runtime. If you have read explainx.ai's guide to agent skills, this is the same pattern — packaged, composable instructions the agent selects from — applied to course construction.

Provider-neutral by design

OpenMAIC does not bundle a model. You bring your own, and the list is long: OpenAI, Azure OpenAI, Anthropic, Amazon Bedrock, Google Gemini, DeepSeek, Qwen, Kimi, MiniMax, Grok (xAI), OpenRouter, Doubao, Tencent Hunyuan, Xiaomi MiMo, GLM (Zhipu), plus Ollama and Lemonade for local models and any OpenAI-compatible endpoint.

Server routes resolve model, media, ASR/TTS, and search configuration provider-neutrally — credentials never reach the browser, any served capability can be force-disabled with a <CAP>_<PREFIX>_ENABLED=false switch, and an unresolved model route fails loudly instead of guessing a vendor. For the workbench specifically you must set MODEL_ROUTES to point maic-agent-driver at a provider-prefixed model with an openai-completions or openai-responses dialect; there is deliberately no fallback.

Running it fully local

For an air-gapped or privacy-constrained setup, the whole stack has a no-API-key path:

That local-first posture lines up with the broader shift explainx.ai tracks in closed-source AI vs. local open-source alternatives — a real classroom generator you can run without sending student material to a hosted API is a meaningful capability for schools and enterprises.

Multi-agent orchestration and the whiteboard

Under the hood, lib/orchestration/ runs a LangGraph state machine — a "director graph" — that manages agent turns and discussions, while an action engine executes 28-plus action types: speech, whiteboard draw/text/shape/chart, spotlight, laser pointer, and more. This is the same LangGraph-based approach seen in harnesses like ByteDance's DeerFlow 2 and Google's Antigravity teamwork framework — see explainx.ai's multi-agent orchestration patterns guide for how these director-and-worker structures are put together and where they tend to break.

The classroom output is genuinely interactive: 3D visualization, process simulations, knowledge mini-games, mind maps, and in-browser coding with instant execution, all responsive down to mobile. The AI teacher can operate that UI directly — highlighting regions, setting conditions, dropping hints — which is closer to explainx.ai's own thesis on learning by making with AI tutors than to a passive video lecture.

Generate a classroom from your chat app

OpenMAIC integrates with OpenClaw, the personal AI assistant covered in explainx.ai's OpenClaw guide. With the skill installed (clawhub install openmaic), you can generate and view classrooms from Feishu, Slack, Discord, Telegram, and 20-plus messaging apps without touching a terminal — hosted mode uses an access code from open.maic.chat, or the skill walks you through a self-hosted clone-and-configure. Every step asks for confirmation first. See the top OpenClaw skills roundup for where this sits among community claws.

Setup

git clone https://github.com/THU-MAIC/OpenMAIC.git
cd OpenMAIC
pnpm install
cp .env.example .env.local
# add at least one provider key, e.g. OPENAI_API_KEY or ANTHROPIC_API_KEY
pnpm dev

Prerequisites are Node.js ≥ 20 and pnpm ≥ 10. For durable server-backed sessions and the workbench, run the PostgreSQL profile:

docker compose --profile server-persistence up --build

and enable the workbench flags (NEXT_PUBLIC_PRO_WORKBENCH_ENABLED=true, OPENMAIC_AGENT_RUNTIME_ENABLED=true) with a matching DATABASE_URL and MODEL_ROUTES.

Honest limitations

• The workbench is not the default, and it is stateful. It needs PostgreSQL, explicit MODEL_ROUTES, and the runtime flags. Enable it without a DATABASE_URL and the session routes just error — by design.
• The shipped persistence auth is not production-grade. The README is blunt: PERSISTENCE_DEV_TOKEN / NEXT_PUBLIC_PERSISTENCE_TOKEN is compiled into the public JavaScript bundle and "provides no confidentiality and no user isolation whatsoever." The server-backed setup as shipped is for localhost or trusted single-user networks; you must replace lib/persistence/server-auth.ts with real authorization before a multi-tenant deployment.
• Video generation is brand new. A commit two hours before this writing made generate_video asynchronous with a placeholder reference — treat video as early.
• AGPL history matters if you forked early. Code before the mid-2026 relicense was AGPL-3.0; the MIT terms apply going forward.
• It is a big system. Next.js 16, LangGraph, a canvas slide editor, an export pipeline, 55-plus React hooks — self-hosting is a real commitment, not a weekend docker run.

Why this matters for builders and educators

OpenMAIC is one of the clearest examples of the agentic classroom idea being shipped as real, MIT-licensed software rather than a research demo. For anyone building AI learning products — the space explainx.ai works in — it is worth studying for three things: the validated-tools approach to letting an agent edit structured content without corrupting it, the provider-neutral capability resolution that fails loudly, and the durable session runtime with leases and crash resume. Those are transferable patterns whether or not you ever run the classroom itself.

Related reading

• What are agent skills? Complete guide
• Multi-agent orchestration patterns guide
• Multi-agent error propagation patterns
• AI tutors and learning by making: Khanmigo, Melo
• ByteDance DeerFlow 2: LangGraph super-agent harness
• Google Antigravity: multi-agent teamwork framework
• YC open-sources QM: company-wide multi-agent harness
• Closed-source AI vs. local open-source alternatives
• What is OpenClaw? Personal AI assistant guide
• Official: OpenMAIC on GitHub · live demo · paper: From MOOC to MAIC, JCST 2026

Repository details — star count, contributor count, license, and the v1.0.0 feature set — reflect the project on August 31, 2026 and will change. Verify the current README and changelog before relying on any specific flag or provider.