OpenMAIC

Why it's ranked here

The engineering discipline stands out more than the pitch. Automated unit and end-to-end tests actually exist, four packages are published to a public registry with automated checks that block a merge if the version number doesn't move, and the history includes a fix for a real deployment bug: a Docker volume ownership problem that silently dropped saved classrooms. It also refuses to lock into one AI vendor, wiring in a dozen-plus model and search providers behind a shared setup file.

What's good

You are not locked to one model vendor: the config accepts keys for OpenAI, Azure, Anthropic, Bedrock, Google, and half a dozen others, and you can run purely from environment variables or a YAML file. Deployment options span a single Vercel-button click, a self-hosted Docker Compose stack, or a hosted service with an access code. The optional video-render container runs on an isolated Docker network with its own egress lockdown, a real security choice given that it executes an unsandboxed browser to composite output.

Tradeoffs

Getting a working install means a sizeable dependency tree and native build tooling for image and canvas libraries, plus a chained build step that compiles several internal packages in sequence before the app itself builds. The optional video-export path needs a separate container that documentation says wants 8 GiB of memory. And since every classroom scene is generated by a language model, you need at least one paid provider API key before anything beyond the interface actually works.

How to use it well

Try the hosted version first with an access code before standing up your own deployment; it skips the environment setup entirely. Educators building a single lesson from an outline or a source document get the most from the one-click generator, while anyone assembling a multi-session course, with materials to upload and edits to make along the way, should reach for the newer chat-driven workbench instead. It plugs into existing chat and IDE assistants too, so a course can start from a Slack message rather than the web app.

Observed

License

MIT

Primary language

TypeScript (Next.js, React), with one small vanilla-JavaScript package for MathML conversion

Packaging

pnpm workspace monorepo with several internal packages (dsl, storage, renderer, importer, editor) published individually to a public npm registry

Deployment surface

Self-hosted via Docker and Docker Compose, one-click deploy to Vercel, or a hosted instance behind an access code

Configuration

Provider keys set via environment variables or a YAML config file

AI provider support

Multiple LLM providers wired in behind a shared interface (OpenAI, Azure OpenAI, Anthropic, Amazon Bedrock, Google, and others)

Testing

Both unit tests (Vitest) and end-to-end tests (Playwright) are present in the repository

Optional components

A separate containerized render service for video export, isolated on its own internal Docker network

Distribution

Ships a standard skill package intended for use from external agent workbenches, not just its own web app