coursecast

One prompt → a full interactive course. Real multi-lesson content with live widgets, quizzes, a custom simulation per lesson, and a mastery dashboard — rendered to self-contained HTML you can open from disk or deploy anywhere with one command.

Prefer zero setup? The hosted version lives at trycoursecast.com — type a topic in the browser and it generates and hosts the course for you.

Install

npm i -g coursecast     # or: npx coursecast …

Requires Node 18+.

Use

# a whole course (~12 lessons + mastery dashboard), from one prompt
coursecast "how DNS works"

# generate and deploy live to your own Vercel
coursecast "intro to nuclear power" --deploy

Open the printed index.html — that's the whole product: no server, no build step, works offline, progress saved in the browser.

Engines

The engine is what writes the content. Pick with --engine:

| Engine | Runs on | Needs | |---|---|---| | claude (default) | your Claude Code CLI | claude on PATH | | codex | your Codex CLI | codex on PATH — run it from anywhere; coursecast passes --skip-git-repo-check | | agy | your Antigravity CLI login | agy on PATH — no API key, but a personal Google account (see below) |

coursecast "learn rust" --engine agy --model gemini-3.5-flash-lite

No engine picks a model for you — pass --model with whatever your key has access to. COURSECAST_RESEARCH=0 skips the web-search research pass.

agy needs a personal Google account. Antigravity is not available on Google Workspace (company) accounts: signing in with one leaves AI credits permanently disabled, and every run dies instantly with Agent execution terminated due to error. Run agy and type /credits to check — if it says credits are not enabled and /settings offers no way to turn them on, you are on a Workspace account. /logout and sign in with a personal address.

All flags

--engine <id>      claude (default) | agy | codex — the CLI you are signed into
--model <id>       model to use, e.g. opus, gpt-5.1-codex-max, gemini-3.5-flash
--effort <level>   reasoning effort: low | medium | high  (claude, codex, agy)
--storage <id>     progress storage: local (default) | cloud
--deploy [host]    build for that host and ship it live: vercel | netlify | cloudflare.
                   Bare --deploy uses whichever CLI you are logged into.
--provider <id>    build for a host WITHOUT shipping yet; deploy it later with
                   --deploy-only. Needs --storage cloud to mean anything, so on its
                   own it is rejected rather than ignored.
--lessons <n>      (course) target number of lessons (default ~12)
--concurrency <n>  (course) parallel lesson generation (default 4)
--deploy-only      deploy an already-generated course (needs --out); never regenerates
--dry-run          with --deploy, print the deploy command without running it
--preview          with --deploy, ship a preview instead of production
--out <dir>        output directory (default ./out/<slug>)

Progress sync

By default a course keeps progress in the browser's localStorage — no server, no account. --storage cloud adds cross-device sync using your host's own database. Nothing is sent to coursecast; the deployed course talks only to its own origin, and the data is yours.

| Host | Storage used | Setup | |---|---|---| | netlify | Netlify Blobs | none — works on deploy | | vercel | Vercel Blob (private) | none — --deploy creates and connects the store for you | | cloudflare | Workers KV | none — --deploy creates the namespace, binds it, and creates the Pages project |

All three are zero-setup: --deploy provisions whatever the host needs.

On Vercel the store is private: reads bypass the CDN cache, so a second device sees the latest progress instead of a copy up to a month old, and nobody's progress sits on a publicly readable URL. On Cloudflare the backend ships as a _worker.js at the root of the deployed folder — Cloudflare ignores a functions/ directory that lives inside the uploaded output, which silently 404s the sync endpoint.

Vercel teams with Deployment Protection: if your team enables SSO protection, every deployment URL asks visitors to log in, so a course you share is not actually readable. Check with vercel project protection; turn it off for a course project with vercel project protection disable --sso, or attach a custom domain.

coursecast "how dns works" --storage cloud --deploy netlify

Naming the host is required with --storage cloud: the three hosts store progress in three different places, so picking one for you would silently decide where your learners' data lives. --deploy <host> says it once and does both jobs — writes that host's sync backend and ships the course there. (A plain --deploy with no host is fine for a local-storage course: guessing a deploy target wrong is immediately obvious, whereas guessing storage wrong is not.)

To build now and ship later, use --provider <host> and then --deploy-only.

The page mints a sync code on first use and shows it in the footer. Enter that code on another device to pick up where you left off. No email, no password, nothing personal stored.

Deploy providers

--deploy ships the output directory with your own vercel, netlify or wrangler (Cloudflare Pages) CLI — whichever you are already logged into, auto-detected — your account, your URL, no middleman. Or skip it and host the files anywhere that serves static HTML. Add a provider via src/providers/<id>.mjs exporting name, isAvailable(), and async deploy({ dir, prod, dryRun }).

Architecture

engine produces a content object (data) → render.mjs turns it into polished interactive HTML (presentation) → provider deploys the folder. Content and presentation are separated, so every lesson looks like one product no matter which engine wrote it. Progress (checklist, quiz, theme) saves per-device in localStorage.

Writing an engine

An engine module exports name, async generateLesson({ topic, slug, model, context }), and — required for whole courses — async generateSyllabus({ topic, lessons, model }). Drop it in src/engines/<id>.mjs and pass --engine <id>. See src/engines/stub.mjs for the minimal shape and src/engines/shared-prompts.mjs for the content contract.

License

MIT © Tomas Sestak