Chat Wrapped
Spotify Wrapped for your WhatsApp and Telegram chats. Drop in an export and get the story of the chat: who talks most, who replies at 2 AM, your emoji language, reply-speed rankings and an award for everyone. An optional AI verdict adds a short personality write-up.
Live: https://wrapped.jashparmar.com · try the built-in demo chat, no export needed.
What you get
A story of 20+ slides and a full stats page:
- Totals and rhythm: messages, words, prime time, a weekday × hour heatmap, a calendar of every day, busiest day, streaks, silences, the longest conversation and the longest message.
- Every person, not just the top few: rank, a personal title from their standout habit (Night Owl, Ghost, Double Texter, Reel Dealer, Lurker…), reply speed, chat bestie, top words and emoji, and a shareable card each.
- Group dynamics: who replies to whom, the dynamic duo, conversation starters and killers, catchphrases, shared links by site, good-morning and good-night champions.
- Your Wrapped: pick which person is you for personal slides; pick a year ("2026 Wrapped") with a year-over-year comparison.
- Play with it: search any word or emoji ("how many times did we say bro?"), a "guess who said it" quiz, CSV export.
Private by design
The interesting engineering constraint: people's chats are about as personal as data gets, so the app is built so that message text never leaves the device.
- Parsing and analysis run in the browser, inside a Web Worker, so a 100k-message chat doesn't freeze the page.
- A strict Content Security Policy (
connect-src 'self') means the page can only talk to its own origin. Even a compromised dependency couldn't send the chat anywhere else, and there are no dependencies anyway.
- The AI sees anonymous numbers only. Names become
P1…P8 before anything is sent ("P1 replies in 3 min, 21% of messages after midnight"), and the server swaps nothing back; the browser does. Top words are opt-in.
- The server re-validates every field against a strict schema (ids must match
P[1-8], words must be single tokens, emojis must be emoji), so the endpoint can't be used as a general-purpose LLM proxy or to smuggle text through.
One tap from WhatsApp
WhatsApp chats are end-to-end encrypted and no app can read them, so instead of a bot the app plugs into the phone's share sheet:
- Android: Chat Wrapped is an installable PWA with a Web Share Target. WhatsApp → Export chat → Chat Wrapped opens the story straight away. The service worker parks the shared file in Cache Storage and the page picks it up, so the chat still never reaches the server (if a share ever hits the server, it redirects without reading the body).
- iPhone: websites can't be share targets on iOS, so the page documents a three-step Shortcut that saves the export and opens the site ready to pick it.
- Back into the group: "Send to the group" opens the share sheet with the summary card and a caption that carries only totals (messages, peak hour, top emoji). Desktop falls back to a
wa.me link. Link previews come from an Open Graph image.
- Offline: the app shell is cached, so it works with Wi-Fi off, which is the easiest way to prove nothing is uploaded.
Analytics without tracking
There's no analytics script on the page (the CSP would block one anyway). The Worker counts page loads itself and the page reports button taps by name (demo, share_group, ...) with sendBeacon. Everything lands in daily counters in D1:
- unique visitors use a salted SHA-256 of IP + user agent + date, so they can't be reversed or linked across days, and the hashes are pruned after 35 days;
- bots and scanners are filtered (user agent plus
Sec-Fetch-Mode: navigate), and link-preview fetches from WhatsApp, LinkedIn and others are counted separately, which shows when the link is being shared;
- a private
/stats/<secret> page (an unguessable link, no login) (server-rendered, no JS) shows visitors per day, the funnel from visit → Wrapped → sent to a group, countries, referrers and AI provider health. Matchday Brain writes to the same database.
Architecture
browser Cloudflare edge
┌──────────────────────────────────┐ ┌────────────────────────────────┐
│ File / zip ─▶ Web Worker │ │ Worker │
│ zip.js (reads only the │ │ /api/insights │
│ chat entry, never │ stats │ validate (schema, sizes) │
│ the media) │ ───────▶ │ edge cache by SHA-256 │
│ parse.js WhatsApp iOS/Android│ (no text)│ rate limit per IP │
│ + Telegram JSON │ │ Gemini → Workers AI fallback │
│ stats.js streaks, replies, │ ◀─────── │ clean + cap model output │
│ heatmap, awards │ verdict │ static assets + CSP headers │
│ Story player · dashboard · PNG │ └────────────────────────────────┘
└──────────────────────────────────┘
| Piece | Notes |
|---|
public/js/parse.js | WhatsApp iOS and Android formats (12/24-hour clocks, bidi marks, multi-line messages, media/deleted/edited markers, system notices) and Telegram JSON. Ambiguous 1/2/24 dates are resolved by picking the reading that keeps the chat in chronological order. |
public/js/zip.js | Reads only the central directory and the chat entry, so a 2 GB export full of photos is never loaded into memory. Inflates with the native DecompressionStream. |
public/js/stats.js | Reply-time medians, conversation starters (6+ hours of silence starts a new conversation), streaks and silences, a weekday × hour heatmap, a gap-filled monthly timeline, top words/emojis and nine deterministic awards. |
public/js/story.js | Full-screen story with auto-advance, tap/hold/keyboard navigation and reduced-motion support. |
public/js/share.js | Renders a 1080×1350 summary card on a canvas and opens the native share sheet when available. |
src/worker.ts | Static assets plus /api/insights: validation, per-IP rate limiting, an edge cache keyed by payload hash (repeat requests return in ~60 ms), and Gemini structured output with a model fallback chain: if a model is overloaded, retired or slow, the next one in GEMINI_MODELS answers instead, and if all of Gemini is busy, Llama 3.3 70B on Cloudflare Workers AI (same account, no extra key) takes over. |
No frameworks, no build step and no npm dependencies.
Develop
npm test # 32 tests: parsers, stats, zip reader, API validation
npm run dev # wrangler dev (needs a workers.dev subdomain for the AI binding)
node scripts/static.js # static server for PWA/share-target testing on :8792
node scripts/compare.js A.zip B.txt # explain why two phones' exports of one chat differ
AI insights use Gemini when GEMINI_API_KEY is set (npx wrangler secret put GEMINI_API_KEY) and fall back to Workers AI, which needs no key.
Deploy
npm run deploy # Cloudflare Workers, custom domain wrapped.jashparmar.com
Not affiliated with WhatsApp, Telegram or Spotify.
