Socratic Algo Coach

A Chrome side-panel coach for competitive programming. It reads the problem beside you, listens to your approach, and helps you reason toward a solution without jumping straight to the answer.

It is open source and uses your own OpenAI API key.

What it does

• Reads problems from LeetCode, DMOJ, Codeforces, USACO, Advent of Code, and many ordinary problem pages.
• Accepts a pasted statement when page extraction is unavailable.
• Uses focused questions, small examples, corrections, and optional diagrams.
• Runs every reply through a separate guard before showing it.
• Offers an opt-in, evidence-based learner profile that you can inspect, correct, export, or erase.
• Never reads or edits the code editor and never submits a solution.

The coaching principles live in MANIFESTO.md. That document is the product contract.

Try it locally

You need Node.js 22+, Chrome 116+, and an OpenAI API key with GPT-6 Astra access.

npm install
npm run build

Then open chrome://extensions, enable Developer mode, choose Load unpacked, and select dist/chrome-mv3. Pin the extension, open a problem page, click the icon, and add your API key in Settings.

Reading the page is local. The first paid request happens when you click Start coaching, after accepting the data-use disclosure.

Contributing

Contributions are welcome. Before changing coaching behavior, read MANIFESTO.md and CONTRIBUTING.md.

Start the development build with:

npm install
npm run dev

Before opening a pull request, run:

npm run format:check
npm run typecheck
npm run lint
npm test
npm run build

No API key is needed for the automated tests; model calls are mocked. Packaged browser coverage is available with:

npx playwright install chromium # once
npm run test:e2e

Good places to contribute include site adapters, extraction fixtures, accessibility, visual explanations, tests, and documentation. Prompt or learner profile changes need focused policy tests; site-adapter changes need a sanitized fixture and a negative case.

Project map

• entrypoints/sidepanel/ — problem extraction and coaching UI
• entrypoints/options/ — API key, privacy, and learner-memory controls
• entrypoints/background.ts — trusted storage, messaging, and model boundary
• src/extraction/ — site adapters and generic page recognition
• src/agent/ and src/prompts/ — analysis, coaching, and response guard
• src/learner/ — local evidence-weighted learner profile
• src/visualization/ — validated diagrams and SVG rendering
• tests/ — unit, policy, security, component, and packaged browser tests

The short version:

problem page → extractor → side panel → background worker → OpenAI
                                                        ↓
                                                  response guard
                                                        ↓
                                                    coach reply

Privacy

There is no application backend and no analytics.

The API key and learner profile are stored in trusted extension-local storage. Browser storage is not a hardware secret store: someone with access to the browser profile may be able to recover the key.

When coaching starts, the relevant problem text and, if enabled, a small learner snapshot are sent to OpenAI. During a conversation, recent messages and private coaching context are sent as well. Sessions use temporary browser storage; learner memory stays local until you erase it. The user pays OpenAI directly. Starting a problem uses one model call, and each normal turn uses two.

Page access is temporary and starts with a user gesture; the extension requests no persistent access to problem sites. Page code never receives the API key, learner profile, or private coaching map.

Read the full PRIVACY.md, or see SECURITY.md for the threat model and reporting guidance.

Limits

Extraction can fail on inaccessible iframes, image-only PDFs, authentication walls, Premium-only statements, and Chrome-restricted pages. Paste the statement when that happens.

The anti-spoiler behavior is a pedagogical guardrail, not a security boundary. A determined user can inspect or modify an open-source extension.

License

MIT