Explore your ChatGPT & Claude history — 100% in your browser.
Drop your export and instantly get full-text search, a usage timeline, per-topic insights, and a clean reader across thousands of chats. Nothing is uploaded, tracked, or stored. Everything runs on your device.
Screenshots use built-in sample data. Your real conversations never leave your browser.
You know you asked your AI that one thing back in March — but the official apps give you no real search and no way to see your own history. chatlens turns your data export into something you can actually explore.
It's a single HTML file with no build step and no dependencies.
- Just open it: double-click
index.html(works offline, straight from disk), or - Serve it:
python3 -m http.serverthen visit the printed URL, or host the one file anywhere static (GitHub Pages, Netlify, an S3 bucket…).
Then load your export, or click Explore with sample data to try it with a synthetic dataset — no export needed.
chatlens accepts three shapes, auto-detected:
- A single
conversations.json— the standard ChatGPT/Claude export. Drop it or click to choose. - A folder of per-conversation files — some ChatGPT exports split each chat into
its own
.json. Drag the whole folder onto the page, or use "select a whole folder." - Multiple files at once — select any number of
.jsonfiles together.
Getting the export
- ChatGPT → Settings → Data controls → Export data → unzip the emailed archive.
- Claude → Settings → Privacy → Export data → unzip.
Both provider formats are detected automatically. chatlens reconstructs each conversation's canonical thread (following the active branch), so regenerated / edited-away branches are excluded from counts — you see the conversation as it stands.
Everything is indexed in-browser in a background Web Worker, so the UI never freezes. Measured on desktop Chrome (M-class laptop):
| Export size | Conversations | Messages | One-time index | Search |
|---|---|---|---|---|
| 30 MB (real export) | 222 | 4,362 | ~1.9 s | 6–15 ms |
| 53 MB (synthetic) | 2,446 | 50,000 | ~7.8 s | ~70 ms |
| 85 MB (synthetic) | 4,857 | 100,000 | ~13 s | ~120 ms |
- Indexing is a one-time cost on load, shown with a progress bar, and scales roughly linearly with total text. Search stays fast (well under a quarter-second) even when a query matches tens of thousands of messages.
- Honest ceiling: past ~100k messages / a few hundred MB, indexing runs into the tens of seconds and memory use climbs (the index keeps a lowercased copy of message text, so peak RAM is a small multiple of the file size). It's built for real personal histories — thousands to ~100k messages — not for indexing a shared multi-account archive. Extremely large exports may be slow or memory-bound on low-RAM devices.
- Full-text search across every message, built as an in-browser inverted index.
- Multi-word queries match messages containing all the words.
"quoted phrases"for exact matches./regex/for regular expressions (e.g./colou?r/).- Filter by you vs the AI, and jump straight from a result into the reader.
- Timeline — messages per month, plus a day-of-week × hour-of-day heatmap.
- Insights — totals, words written, busiest day, peak hour, models used, longest conversations, and what you talk about most.
- Reader — a clean serif reading view of any conversation with lightweight Markdown (code blocks, lists, bold, links) and per-message model labels.
- Light & dark themes (paper by default; a warm "reading at night" dark mode).
- No backend, no network calls. Open your browser's Network tab, or disconnect from the internet — it works the same. Reload the tab and all data is gone.
- Parsing and indexing run in a Web Worker; the main thread only renders. If a browser
blocks Blob-URL workers (some
file://setups), it transparently falls back to a same-interface main-thread parser. - The web client is a single vanilla HTML/CSS/JS file — easy to audit end to end.
See PRIVACY.md for a 2-minute checklist to verify the local-only behavior yourself (offline test, Network tab, reload-disposal, source grep).
No cloud, no LLM calls, no "continue this chat." chatlens reads and understands your existing history; it doesn't talk to any model.
MIT