OpenMAIC's agent workbench elevates AI classrooms beyond glorified slide generators

OpenMAIC turns a topic prompt or uploaded document into a full interactive classroom, complete with AI teachers who speak, draw on a whiteboard, and field questions in real time. That's the core proposition, and it's been iterating fast: the project went from a basic one-click generator to a full agent workbench between March and August 2026, with the v1.0.0 release adding durable, server-backed course-building sessions that survive restarts and accept follow-up steering mid-run.

The architecture is more considered than the marketing framing suggests. Multi-agent orchestration runs through LangGraph, and the action engine handles over 28 distinct action types — speech, whiteboard drawing, spotlight effects, laser pointer animations. The classroom isn't just a slide deck with a chatbot bolted on; it's a state machine driving playback through distinct modes: standard lecture, roundtable debate, Q&A, and project-based learning where you pick a role and collaborate with agents on structured deliverables.

The v1.0.0 agent workbench is the most interesting addition. Rather than generating a course in one opaque pass, it exposes an explicit tool surface — the agent plans curricula, patches individual scenes atomically, searches extracted material text, generates media, and imports .pptx files with layout preserved. Twenty built-in skills cover teaching styles from deep research to vocational training. Sessions are PostgreSQL-backed with leases, heartbeats, and crash resume, and database-maintained revision counters mean the workbench only refetches scenes that actually changed.

Provider neutrality is taken seriously. The README lists over a dozen LLM providers — OpenAI, Anthropic, Google, Amazon Bedrock, DeepSeek, Qwen, MiniMax, GLM, Ollama, and more — plus separate provider slots for TTS, ASR, image generation, video generation, and web search. Local inference is a first-class option through Lemonade and FunASR, with no API key required. Credentials stay server-side; the browser never sees them.

The persistence story is honest about its limits. The development token compiled into the browser bundle provides no real user isolation — the README says so plainly, and it tells you exactly which file to replace before any production deployment. Asset collection, S3 egress, and horizontal scaling are all addressed with specific configuration knobs rather than hand-waving.

The interactive mode — 3D visualization, physics simulations, in-browser coding, knowledge games, mind maps — is genuinely differentiated from a typical slide generator. The AI teacher can operate the interactive UI directly, highlighting areas and setting conditions during a lesson. Export covers editable .pptx, self-contained HTML with inlined assets for offline use, and full classroom ZIP.

The codebase is a Next.js 16 / React 19 / TypeScript 5 monorepo with a published @openmaic/* SDK family covering the DSL, renderer, editor, importer, generation contracts, and storage primitives. The project is backed by a JCST paper from Tsinghua and relicensed from AGPL to MIT in June 2026, which removes the main friction for commercial deployment.