transcrub

A local, private co-reading scrubber for your AI chat transcripts. Merge transcripts from multiple sources into one chronological timeline, scroll through them with full context, see extended-thinking blocks, and leave margin notes ("marks") — yours and the model's — pinned to individual turns.

Runs entirely on 127.0.0.1. Your transcripts never leave your machine.

Walkthrough on a synthetic Lightcell example: open the conversation, smart-scroll through the reasoning (thinking blocks in purple), jump to a margin mark, search, and pin a note. A few easter eggs are hidden in the marks — mp4.

See it in your browser

Live demo (no install): https://daniellefong.github.io/transcrub/ — loads synthetic example data.

Run locally (zero dependencies — just Python 3):

git clone https://github.com/DanielleFong/transcrub && cd transcrub
python3 run.py        # builds your index if sources.json exists, else example; opens your browser

Use it on your own transcripts

1. cp sources.example.json sources.json and point each entry at your data:
   • claude-code — Claude Code .jsonl sessions (~/.claude/projects)
   • claude-json — conversation JSON with a chat_messages[] array (export / API shape; keeps thinking)
   • text — plain-text transcripts (best-effort)
2. python3 scan.py → builds data/index.json
3. ./serve.sh

To export your own Claude conversations with extended thinking (which the share/export drops), see tools/.

What you can do

• Corpora are colored bands; filter them with the chips. operator-own vs testimony (other people's transcripts) are tagged as distinct evidentiary classes and never blended.
• thinking only isolates turns that carry reasoning blocks.
• Scroll the feed for context (preceding + current + following); the current turn advances when you reach its bottom. Arrows / filmstrip-click / marks smooth-scroll to a turn.
• Marks: pin a note to any turn (author = you or the model). Saved to your browser, exportable to markdown. Seed marks live in the SEEDS array in coread.html.

Privacy

data/ (except the synthetic example), sources.json, export/, uploads/, and tools/output/ are gitignored. The viewer binds to localhost only. Don't deploy a build that contains real transcripts to a public URL without your own auth in front of it.

License

MIT