Twenty-four skills turn a prompt into an editable classroom
OpenMAIC ships 24 built-in skills for planning curricula, researching material, shaping slides, importing PowerPoint files, and creating different teaching formats. A prompt or uploaded document can become slides, quizzes, interactive HTML, project work, narration, and a cast of AI teachers or classmates. The Pro workbench lets an agent revise individual scenes without rebuilding the lesson.
That is a larger job than making a deck. OpenMAIC stores durable agent sessions, lets a run resume after a worker restart, and supports follow-up instructions while generation continues. The classroom can be exported as editable PPTX, interactive HTML, or a course archive. Its interface covers 12 locales across 11 languages. You still own the factual review, because the README describes generation and editing tools rather than an automatic check of lesson accuracy.
Node.js 22.19, pnpm 10, and PostgreSQL 16 are the floor
A local source setup starts with Node.js 22.19 or newer, pnpm 10 or newer, and PostgreSQL 16. The server refuses to start without DATABASE_URL. You also configure at least one language-model provider in openmaic.yml or through the compatible environment settings. Model slots can route the outline, course content, agent, image work, speech, search, and video to different services. That control is useful, but each extra route adds another credential or endpoint to test.
Local options reduce the need to send material to a hosted vendor, though they do not remove the operating work. Lemonade can cover language, image, speech, and transcription routes. FunASR handles local speech recognition, while local audio and video extraction needs ffmpeg and ffprobe on the path. The base app skips that extractor if those programs are missing.
What happened when we ran it
Our sandbox installed commit f287542 in 176 seconds, adding 2,546 pnpm packages and using 2,958 MB on disk. The checkout itself held 3,100 files, about 616,786 lines of source, and occupied 116.9 MB. OpenMAIC then built successfully in 162 seconds. The Debian container had 3 CPUs, 8 GB of RAM, no secrets, and no elevated privileges. These figures cover repository mechanics, not lesson quality or generation speed.
The test command failed after 371 seconds. Vitest reported 8,848 passed, 1 failed, and 98 skipped across 8,956 tests. The failing file was tests/video-export/eslint-boundary.test.ts, where 1 of its 8 tests exceeded the 5,000 ms limit. The log also recorded 1 error. It does not show why the boundary test missed its deadline, so blaming the container, the test, or the code would go beyond the evidence.
The same checkout contained 6 CI workflow files, a Dockerfile, Compose configuration, a tests directory, and monorepo workspaces. That engineering surface is meaningful, as are 8,848 passing tests. The single timeout still makes the complete result red. A team that blocks releases on a clean suite should reproduce that file on its own runner and decide from the new log, rather than treating the passing build as a substitute.
The 2-container Docker default is a personal install
Docker Compose starts exactly 2 containers: the application and PostgreSQL. It binds the app to 127.0.0.1:3000, keeps the database off the host network, and maps every request to one owner. Named volumes retain courses, generated media, and runtime sessions. Those defaults make sense for one person on one machine, but they are a poor match for a school server exposed to several unrelated users.
Network access requires deliberate changes. The README tells operators to set a long ACCESS_CODE, replace the initial PostgreSQL password, and publish on 0.0.0.0. That access code is a shared site password, not a student directory or role system. OpenMAIC has no built-in gateway authenticator. A public deployment needs its own identity layer and authorization rules before separate learners can safely share the service. Moodle starts much closer to that requirement.
MP4 export adds Chromium and FFmpeg to a separate service
Video export is an optional Compose profile, not a feature of the base 2-container stack. The browser creates a Hyperframes project, then an isolated Node.js 22 render service uses Chromium and FFmpeg to turn it into an MP4. Without that service, OpenMAIC downloads a project ZIP for local rendering. Operators must plan its concurrency and resource use separately from the web application and database.
The agent workbench also needs an explicit production choice. Two flags enable its public entry point and server runtime, and its model slot must support tool calls. PostgreSQL stores its sessions, leases, heartbeats, materials, and user-authored skills. None of this is hidden in the documentation. It does mean the most interesting mode is not what appears after a bare install with one model key.
Open issue 1772 documents a v1.1.2 narration recovery gap
Open issue #1772 reports 65 speech lines being generated with no audio or warning when a self-hosted v1.1.2 instance had no usable server-side TTS binding. The same report says a TTS failure during generation left completion flags unset and the editor disabled until the operator changed the database manually. This is a user report, not a failure from our sandbox, but it concerns a central classroom feature and names a concrete recovery problem.
Six CI workflows and an October 2 push show current maintenance
GitHub showed 39,827 stars, 156 open issues, and 24 open pull requests on October 3, 2026. The repository was pushed on October 2, and several issues and pull requests were updated that day. Release v1.1.2 arrived on September 28 with stricter validation and redirect handling for caller-selected provider URLs. Those dates show active security and maintenance work alongside a sizable unresolved queue.
OpenMAIC is the best fit when the classroom format itself matters: sequenced scenes, live interaction, narration, editable course state, and several export paths. Our 162-second build proved the pinned commit compiles, while the 371-second test run did not finish cleanly. Give it a sandbox, break a provider on purpose, and verify recovery before a learner depends on the result.

