OpenMAIC

The repository is free, and the README states that the MIT license permits commercial use free of charge. The getting-started guide still asks for at least one LLM provider before the app will generate anything. Cloud providers such as OpenAI or Anthropic usually need a paid API key, while a local Ollama or Lemonade server needs no key.

The README recommends turning on every modality, including generated illustrations, narration, video clips and web search. Each of those can be a separate provider with its own bill. Budget per lesson from your provider dashboards, because the project publishes no cost estimate for a generated classroom.

No. The demo at open.maic.chat is the only hosted address the README links. Its Login button opens a Chinese-language MAIC account page with password or phone-code sign-in, and the v1.1.0 notice on the demo says you can connect your own Token Plan when your credits run out.

Credit amounts and hosted terms are not published in English. Self-hosting avoids that dependency but moves the work to you: Node.js 22.19 or later, pnpm, provider keys and a server. The OpenClaw skill can use either route, with a hosted access code or a local checkout.

Version 0.3.0 relicensed the project from AGPL-3.0 to MIT, and the repository LICENSE file now carries the MIT text. Under MIT you may modify the code and ship it inside a paid product without publishing your changes, which the AGPL would have required for a network service.

Two bundled packages are outside the root license. The README lists mathml2omml under LGPL-3.0-or-later and a third-party pptxgenjs copy under MIT, so redistributing the whole repository means honoring each package license. Pages elsewhere that still describe OpenMAIC as AGPL predate the change.

Every classroom can be exported in three forms, so the result is not locked inside the app:

• Editable .pptx slides with images, charts and LaTeX formulas.
• Self-contained interactive HTML pages with the simulations.
• A classroom ZIP with course structure and media; exports inline external libraries and fonts so the course can play offline after import.
• MP4 video, only when the separate render-service container is running; otherwise you get a project ZIP to render locally.

The README describes the default as fail-open. When ACCESS_CODE is unset, every matched route, including the API, is reachable, so anyone with the URL can spend your model credits. Set a random code of at least 16 characters, and add TRUST_PROXY_HEADERS behind a trusted proxy to get its 10-attempts-per-minute throttle.

With server persistence on, anyone who can reach the endpoint and knows a course ID can read that course, and learners without accounts are anonymous owners tied to a 30-day cookie. The project shipped four security releases in September 2026 alone, so plan to update a shared deployment regularly.

Several advertised features are opt-in. The Pro agent workbench needs PostgreSQL, two feature flags and a model route; MP4 export needs the render-service container; complex-table parsing can use MinerU; and local speech recognition runs through FunASR.

Voice cloning through the optional VoxCPM2 adapter stores an uploaded reference clip in the browser and sends it to your voice server on each synthesis. Clone only voices whose owners have agreed, and treat recordings of students as personal data.

1. Record the release, model, parser and any media providers enabled.
2. Check the generated outline against the PDF headings before scenes are built.
3. Play the classroom and note factual errors in narration, quiz answers and the simulation.
4. Export .pptx and the classroom ZIP, then open both on another machine.
5. Read the provider dashboard for the tokens and cost of the run.

Accept a lesson whose formulas and table survive the .pptx export, whose quiz answers match the PDF and whose cost is recorded. Reject missing formulas, a quiz with wrong keys or an export that does not open. No lesson quality or cost is reported until the run is done.

If you only need a slide deck rather than a taught classroom, compare Gamma and Dokie AI. To study your own sources with cited answers instead, see NotebookLM.