OpenTutor

Compounding deliberate curiosity.
A personal tutor for whatever you're curious about: on the web, on Telegram, or inside your AI agent.

---

opentutor.mp4

_{Music: "Equatorial Complex" Kevin MacLeod (incompetech.com), licensed under Creative Commons: By Attribution 4.0. Excerpted, faded and synchronized with this video.}

OpenTutor teaches one short lesson at a time. It asks before it explains, works on what you keep getting wrong, and brings ideas back days later in a new form.

It ships with 293 courses, from game theory to bread chemistry, and can build a new one for any topic you name.

Try it

opentutor-gg.vercel.app: browse every course and answer an example question, no account needed.

Create an account for 3 free lessons. After that, connect your own OpenRouter account to keep going; you pay OpenRouter directly for what you use. You can delete your account and data at any time (privacy).

Run it on your computer

You need Node 22 or later and a model to teach with: an API key for an AI provider, a local model through Ollama, or Claude Code.

git clone https://github.com/LEARNableLabs/opentutor
cd opentutor && npm install

export OPENTUTOR_LLM=openrouter OPENROUTER_API_KEY=sk-or-...
npm run web

Open http://localhost:3000/learn.html. Your progress is saved on your computer, in workspace/. Lesson text goes to the AI provider you chose.

• Set the variables in your shell. npm run web does not read .env. To use a file, run node --env-file=.env scripts/web/server.js instead.
• Other providers: use claude-sdk with ANTHROPIC_API_KEY, openai with OPENAI_API_KEY, ollama for local models, or cli for Claude Code with no key.

On Telegram: put TELEGRAM_BOT_TOKEN and the same provider settings in .env (the bot reads it), then run npm run bot. See Telegram on a separate host.

In your AI agent: run npx skills add LEARNableLabs/opentutor, then follow the guide for Claude Code, Codex, Claude Web, Hermes, OpenClaw, NemoClaw or NanoClaw.

How it teaches

• It asks before it explains. A lesson is a short conversation: a question about what you already think, a follow-up, then a real situation where you explain why.
• It comes back to what you missed. On the website and Telegram, a rules-based check (no AI) runs after each lesson and notes the concepts you're still unsure of. One that is still shaky three lessons after it was flagged comes back as a retest at the start of a lesson; after five, the next lesson is a review of it.
• On Telegram it adapts further. The same check also raises or lowers the difficulty and changes the kind of question, and each idea gets its own review schedule that stretches as you remember it.

Lessons take a few minutes. On Telegram you can say "skip" at any point. In an AI agent, the agent itself follows the method in the skill's instructions. docs/methodology.md explains the method, the research behind it, and where it falls short.

Where the courses come from

The 293 shipped courses live in skills/tutor/domains/, 5 to 40 lessons each; the website lists them all. They are AI-generated and have not all been rebuilt or approved by the current pipeline. Their citations and resource lists have not been systematically verified and can contain invented or dead links. A working link alone does not establish that it identifies the cited work or supports the lesson's claims.

For a new topic, OpenTutor searches public sources (arXiv, Semantic Scholar, OpenAlex, Wikipedia, course syllabi, YouTube and GitHub), plans a course, and has a second AI pass review it, with up to three build/review rounds (the initial draft and at most two revisions). Sources can be unavailable, and a course can remain unapproved after the third round. On the website and Telegram, five starter lessons are ready straight away while the full course is built. On the hosted site, new topics need your own OpenRouter account.

Host it for others

• A website anyone can sign up to (Vercel and Supabase): follow docs/self-deploy.md, and run the four migrations in supabase/migrations/ before deploying. Sign-up opens automatically once Supabase is configured. The free trial runs on your key, capped at 300 model calls a day (OPENTUTOR_TRIAL_CALLS_PER_DAY); see free trial and students' own keys.
• A small group on one server: run npm run host. The first run writes a learner password and a different admin password to .env, readable only by you (on Windows, which ignores that, restrict the file yourself), and says where they are without printing them; later runs reuse them. The learner password opens the shared workspace; the admin one opens /admin.html, where you add students. Each student gets a one-time access token to enter at /login.html ("Have an existing access token?"), and their own progress. Signing in and every lesson send a password or a token, so serve it over HTTPS: put a web server such as Caddy or nginx in front. Without a password, OpenTutor only answers the computer it runs on.

Documentation

Read this	For
docs/methodology.md	How it teaches, and the research behind it
docs/architecture.md	How it works: the agents, the lesson flow, the course files
docs/self-deploy.md, docs/deployment.md	Hosting on Vercel and Supabase, Telegram, local development
CLAUDE.md, AGENTS.md	Working on the code: every change gets an issue and a PR, reviewed by CodeRabbit and Codex

License

MIT. See LICENSE.