A reader that zooms papers from a five-word map to every word

What it does

TextFold takes almost any text source — arXiv links, DOIs, web pages, PDFs, Word/PowerPoint/Excel files, EPUBs, images, GitHub READMEs — and renders it as a zoomable document with five levels of detail. At level 5 the whole paper is a map of section cards; at level 1 you’re reading the original text. Levels 2 keeps the author’s own key sentences verbatim; levels 3–5 are AI-generated summaries that rest on those sentences, so the abstractions stay anchored to what was actually written.

The interesting bit

The folding isn’t a separate summary view — as you zoom, the surviving words slide to their new positions and the rest fade out, so you can stop halfway or open a single paragraph. It also folds agent output: a Claude Code mod and a Pi extension apply the same five levels to replies, one keypress per level, with no second model involved.

Key highlights

• Local-first: runs on your machine, uses a model you already have (Claude Code, a ChatGPT plan, OpenRouter, or any OpenAI-compatible endpoint including Ollama), no account to create.
• Ask AI about what you’re reading; answers come structured with each point citing the paragraph it rests on.
• Highlights and margin notes persist as you zoom out, and save with the page.
• Sentence-by-sentence translation into many languages, formulas typeset.
• Integrates with Zotero (reads your library via the local app, nothing sent to Zotero’s servers) and Obsidian via a community bridge plugin; python3 -m textfold get feeds documents to agents one level at a time.

Caveats

• Scanned PDFs need the optional MinerU PDF engine, a ~3.4 GB download.
• DOI-based papers need an open full text on Europe PMC; arXiv papers without HTML fall back to PDF reading.
• The ChatGPT-plan bridge needs Node.js, and WorkBuddy support is labeled experimental.

Verdict

Researchers and anyone drowning in papers or verbose agent output should try the demo — the semantic-zoom reading experience is genuinely different. If you only ever read short documents, it’s a lot of machinery for a problem you don’t have.