Signal Lens

The Problem

Months of chats with Claude, ChatGPT and half a dozen coding assistants pile up, and nobody actually knows what’s in them — how much they’re costing in API-equivalent terms, whether a chat leaked a password, which weeks were genuinely productive versus just long. Every one of those tools’ own export is a raw JSON or Markdown dump; nothing reads it back and tells you anything about it.

What it does

Imports exports from Claude, ChatGPT, Google Gemini, Microsoft 365 Copilot, GitHub Copilot, Cursor and Claude Code — or a generic CSV/JSON/Markdown transcript — and turns them into activity trends with a forecast, distinctive topics via TF-IDF, a “quality” read built from rule signals like pushback and self-correction rather than another model’s opinion, a Luhn-verified sensitive-data scan for emails/keys/passwords, and an API-equivalent cost breakdown that accounts for the “long-chat tax” of re-sending history every turn. Every chart mark opens the conversations behind it.

How it is built

Classic scripts on one namespace, no bundler and no build step — it runs from a static host or straight off disk. Each source has its own adapter that normalizes an export into one shared format, stored in the browser’s IndexedDB, so no conversation ever leaves the machine it was imported on — a hosted copy of the site never receives anyone’s data either way.

A simple tech stack workflow to understand

Export a .zip / .json / .md from Claude, ChatGPT, Copilot, Cursor, etc.
   → drop the file in the browser
   → an adapter detects the source and normalizes it
   → stored in IndexedDB (never uploaded)
   → analytics: TF-IDF topics, regex quality signals, sensitive-data scan, cost model
   → charts, drill-downs and notes, all client-side

Where it stands

Live as a static deploy on Vercel — the same build runs unmodified on any static host. Source on GitHub.