Generate quizzes with an AI that runs entirely in your browser, and challenge friends with a link. No server, no accounts, no API keys, no cloud bills.
EasyQZ is a complete rewrite of the original Flutter + server app into the AgentHerd pattern — local WebLLM/WebGPU inference, IndexedDB storage, and self-contained invite links for social play. Everything happens in the tab. Nothing is uploaded.
🔗 Live: easyqz.online · 🧩 Static-hosted on GitHub Pages · 🗄️ Original Flutter/server version preserved at tag
flutter-legacy
- 🤖 AI quiz generation in the browser — a local model writes the questions on your device via WebGPU. No key, no backend, no per-token cost.
- 🎛️ Four ways to make a quiz — a topic, pasted text/article, a category (random subtopic), or an offline Quick Quiz.
- ⚡ Quick Quiz (works offline / no GPU) — multiple-choice questions built from a bundled bank of 500, so you can play even without WebGPU or a downloaded model.
- 🔗 Challenge a friend with a link — the whole quiz and your score are packed into the URL. Your friend opens it, takes the same quiz, and sees who won. No one needs to be online at the same time.
- 🏆 Local leaderboard & history — every score that passes through your browser (yours + incoming challenges) is ranked and saved locally.
- 🔒 Private by design — your text, quizzes, and scores never leave your machine.
Your browser tab = the whole app
┌────────────────────────────────────────────────┐
│ WebLLM (WebGPU) → generates the quiz │
│ IndexedDB → identity, history, scores │
│ URL-hash link → async challenge a friend │
└────────────────────────────────────────────────┘
No HTTP / SSE / WebSocket. No backend.
The only network traffic is the one-time model download (cached afterward) from the public MLC/WebLLM CDN. Quiz content, answers, scores, and your identity stay in the browser. Challenges travel inside the link — nothing is stored on any server.
| Mode | Needs a model? | What it does |
|---|---|---|
| ✍️ Topic | yes | Type any topic — the model writes questions about it |
| 📄 From text | yes | Paste an article or notes — questions test that material |
| 🎲 Category | yes | Pick a category; a random subtopic keeps every quiz fresh |
| ⚡ Quick Quiz | no | Instant MCQs from the bundled 500-question bank |
Categories: Science · Math · History · Geography · Technology · Sports · General Knowledge.
- Finish any quiz → tap 🔗 Challenge a friend.
- EasyQZ compresses the quiz + your score into a link:
…/#challenge=<token>(deflate-raw + URL-safe base64 — the same packing AgentHerd uses for WebRTC tokens). - Send the link via any chat app, email, or SMS.
- Your friend opens it, takes the same quiz locally, and the result screen shows You vs Them. They can re-challenge back with a fresh link carrying both scores.
Because the challenge lives entirely in the link, no one has to be online at the same time and there's nothing to host.
| Model | Size | Notes |
|---|---|---|
| Llama 3.2 · 1B | ~0.9 GB | Fastest — good for quick sessions |
| Qwen 2.5 · 1.5B | ~1.1 GB | Efficient, multilingual capable |
| Gemma 2 · 2B | ~1.5 GB | Balanced quality/speed |
| Llama 3.2 · 3B | ~2.3 GB | Better, more reliable quizzes |
| Phi-3.5 Mini | ~2.2 GB | Strong reasoning |
Larger models follow the JSON quiz format more reliably; 1B models are fastest but occasionally need a retry (handled automatically).
- No global leaderboard or trending. With no shared backend there's no global store — leaderboards are local plus whoever's challenge links you've played.
- No "quiz this URL" fetch. A pure-browser app can't fetch arbitrary pages (CORS), so use From text and paste the content. It's captured into the quiz, so the challenge link stays self-contained.
- No login. Identity is a local, self-asserted name + avatar stored in your browser. Clearing site data resets it.
npm install
npm run dev # http://localhost:3000Requirements: a WebGPU browser (Chrome/Edge 113+) for the AI modes; a GPU helps (integrated is fine for 1B). Quick Quiz works everywhere, including browsers without WebGPU. The first AI quiz downloads model weights (~1 GB) and caches them, so later quizzes start instantly.
npm run build # → dist/ (static files only)
npm run preview # serve the production build locallyPushing to main builds and deploys to GitHub Pages via
.github/workflows/deploy.yml. To activate:
repo → Settings → Pages → Source → GitHub Actions. For the custom domain, add a
CNAME file containing easyqz.online (and point DNS at GitHub Pages).
The Vite base is relative, so the same build works on a custom domain and on a
GitHub Pages project path (/easyqzm/) without reconfiguration.
index.html · style.css · challenge-bank.js (bundled 500-question bank)
src/
main.js app bootstrap, hash routing, all screens
llm.js WebLLM/WebGPU wrapper + quiz prompts → structured JSON
quiz.js AI + offline generation, scoring, persistence, leaderboard
db.js IndexedDB stores: profile · attempts · quizzes · challenges
identity.js local self-asserted identity (name + avatar)
share.js encode/decode a challenge ⇄ URL hash
.github/workflows/deploy.yml GitHub Pages
| What | Why |
|---|---|
| WebLLM | Runs LLMs in the browser via WebGPU — no API key, no cloud |
| IndexedDB | Local persistence for identity, history, and scores |
| CompressionStream | Packs a full quiz + score into a shareable URL hash |
| Vite | Build tooling and dev server |
| GitHub Pages | Static hosting — serves only HTML/JS/CSS, no server-side compute |
The trust boundary is possession of a challenge link — anyone with it can take that quiz, and the challenger's name/score in it are self-asserted (no cryptographic identity). It has the threat model of a group chat, not a platform. Everything else is local to your browser. This mirrors AgentHerd's design: safe among trusted peers, appropriate for a client-only app.
The original Flutter client and its Hugging Face Space server (HTTP/SSE/WebSocket +
Google-JWT auth) live on at tag flutter-legacy:
git checkout flutter-legacy # restore the old app
git checkout main # back to the client-only rewriteMIT — built by Vishal Mysore.