Blocks are laid out in columns by the client, never by the model. Pan, pinch or ctrl-scroll to zoom, hold space to drag anywhere. KaTeX renders the maths.
Source available · runs on your machine
A tutor at a whiteboard, on your own machine.
OpenStudy talks you through any topic while it writes on a zoomable board in handwriting, sketches diagrams, circles what it means, asks you questions and adapts to your answers. You bring the subscriptions and keys. Nothing leaves your computer that you don't send.
- 1 hard requirement: the
claudeCLI - 0 keys needed for the demo
- ~$1 an hour with voice + pictures
talks✦writes in handwriting✦sketches diagrams✦shows pictures✦asks questions✦adapts✦listens✦remembers the board✦
✦✦✦✦✦✦✦✦
Live demo · no model needed
Watch the tutor write on the board
The tutor doesn't place pixels. It streams a small line protocol, and the board turns each line into handwriting, sketches and pen marks. Press play, or click any line to jump to it.
Press play to start the lesson.
How it teaches
A real lesson, not a wall of text
Every lesson follows the same rhythm a good tutor uses. Click through the steps.
It asks before it teaches
Type a topic, attach notes, or point it at a folder of slides. The tutor greets you, puts a title on the board and asks one or two real questions: what you want to be able to do, and what you already know.
- Attach PDFs and text files, read in the browser
- Pick Sonnet, Opus or Fable for the live tutor
- Or try the scripted demo with no model at all
A plan, made while you talk
While the intake is still going, a separate planner call on Fable lays out the lesson as keypoints, and generates pictures ahead of time so they're ready when they're needed. The first teaching turn starts on a provisional plan if the real one is still coming.
- Keypoints track progress in the corner of the board
- Pictures from Gemini, gpt-image-1 or Flux, drawn ahead
- Structured output, one call per lesson
Keypoints 2 / 4
- What reinforcement learning is
- Three ways to learn a skill
- Learning to ride a bike
- From reward to learning
It writes and talks at the same time
Each turn is a small step: two to five spoken lines and a few board blocks. Text streams onto the board in handwriting as the tutor speaks, diagrams are re-drawn with a hand-sketched look, and it tells you where to look.
- Handwriting, KaTeX maths, tables
- Sketched SVG diagrams and generated pictures
- The camera follows the pen, or roam freely
It checks you actually got it
At the end of a section the tutor asks a question that tests understanding, never "does that make sense?". It can offer quick choices you click, or you just answer out loud.
- Quick-check options or free answers
- Speak or type, whichever you like
- Interrupt any time; talking over it stops it
If a robot vacuum bumps a wall and loses a point, which part of the loop is the lost point?
Try it, click an answer.
It adapts, and keeps the board
Wrong answer? It goes back with another angle. Breezing through? It skips ahead. Every lesson is saved, so you can reopen a board from the start page tomorrow and keep talking to the same tutor session.
- Saved in SQLite, no native dependencies
- Resume the same session by id
- Print any board or save it as a PDF
Features
Everything a whiteboard can do, and a few things it can't
Tap a card to open it up.
Annotations are measured from text ranges or picture regions, so the circle lands exactly on what the tutor is talking about, even on pictures and pages of your PDFs.
Charts, flows and geometry come out of the model as plain SVG and get the sketchy treatment on the board. No keys required.
Gemini 2.5 Flash Image by default, or gpt-image-1, or Flux. The tutor always tells you what to look at in a picture and marks it.
With a Deepgram key the tutor speaks with Aura-2 and listens with Nova, streamed through your local tutor server so audio starts in a few hundred milliseconds. Without one, the browser's own speech is used. The board keeps step with the voice either way.
Pages come onto the board as pictures with word boxes from the PDF text layer, or from local OCR (tesseract, no network) for scans and screenshots. The tutor marks on your actual words.
A figure library built with the Anatomy skill, shown in sandboxed frames. Animated @motion diagrams for movement, and @build_figure jobs that build a brand new model in the background.
Boards live in a local SQLite file. Reopen one from the start page and pick up the conversation where you left it, or print it and keep the notes.
Screenshots
Take the tour
Real captures from the scripted demo lesson. Click any image to view it full size.
Drag the board
The finished RL 101 board, at full size. Drag, swipe or use the buttons.
100%
✋ drag me
Setup
Build your setup
Only the claude CLI is required; the tutor and planner bill to the Claude plan you're logged in with.
Everything else is optional. Pick what you want and copy your .env.
rough cost per hour of lesson
~$0.95
on top of your Claude plan
-
1
Sign in to Claude
Install Node 22, pnpm and the
claudeCLI. Runclaudeonce and sign in. -
2
Clone and install
git clone https://github.com/delsqrdpsi/openstudy && cd openstudy && pnpm install -
3
Add keys and start
cp .env.example .env && pnpm devWeb on
localhost:3000, tutor server on:8787.
Under the hood
Small pieces, plainly put together
apps/web
The board
Next.js. Blocks, column layout, pan and zoom, handwriting, rough.js sketches, annotations, voice.
apps/tutor
The tutor server
Websocket sessions, the claude CLI headless per turn, the planner, image and voice adapters. Holds your keys.
packages/shared
The contract
Block schema, lesson events, the streaming protocol parser, the board reducer and layout.
FAQ
Questions, answered
Do I need an API key to try it?
No. The scripted demo lesson on the start page runs with no model at all. For live lessons you need the claude CLI signed in to your Claude plan; the tutor and planner bill there, not to an API key.
What does it cost to run?
The Claude plan covers the tutor. A generated picture is about a cent or two, and an hour of Deepgram voice is under a dollar. A small ring at the top of each lesson shows the running cost live.
Does my data leave my computer?
Keys stay in .env on your machine. The browser only talks to your local tutor server, which calls the providers you've configured. Boards, generated images and prepared pages are stored locally under ./data.
Which browsers work?
Any modern browser for the board. With a Deepgram key, voice works everywhere too. Without one, the microphone falls back to the browser's speech recognition, which needs Chrome or Edge.
Can it teach from my own course material?
Yes. Point it at a folder: PDFs, decks (pptx, odp, key, converted with LibreOffice when installed), images and text. Pages land on the board and the tutor annotates your actual words.
Can I use it commercially?
OpenStudy is source-available under the PolyForm Noncommercial License 1.0.0: free for personal study, research, hobby projects, teaching, charities, schools and public institutions. For a commercial licence, contact @delsqrdpsi.
What do you want to learn today?
Clone it, sign in, ask. The board is waiting.