https://github.com/user-attachments/assets/721a646c-04e7-4618-a527-b43dde64d088
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
License
MIT. See LICENSE.
