TextFold folds a long paper or AI reply into five zoom levels

Your coding agent writes more each day than you can read in one sitting. TextFold, an open-source reader released this month by Xia Chen, shows any long text at five levels of detail and lets you zoom from a one-sentence spine down to every word, on your own computer.

The project is on GitHub under Apache 2.0 and had 401 stars on 20 October, sixteen days after it was created. It was formerly called PaperFold. The five levels, named in the README, run from Topic (a few words per section) to Takeaway (one line per section), Key (the point of each paragraph and the reason for it, written by a model), Brief (one to three verbatim sentences from the author) and Full (the paragraph as written). The levels marked AI rest on the author's own Brief sentences, which the project cites below each AI line.

What it reads

TextFold takes a paper link from arXiv, PubMed Central, bioRxiv, medRxiv or Wikipedia, a GitHub README, or any other web page and strips the menus. It opens a PDF, a Word, PowerPoint, Excel or CSV file, an EPUB, a saved HTML page, Markdown, plain text, or a picture read through its text. A Zotero library opens straight from the Zotero app the reader already runs, with no account and nothing sent to Zotero's servers. A scanned PDF or a picture needs the optional MinerU engine, which the installer offers at about 3.4 GB.

Three ways to run it

A desktop app for macOS on Apple silicon or Intel, and for Windows, carries Python inside. On any other system, python3 -m pip install textfold and python3 -m textfold serve opens the reader on localhost port 3017, with Python 3.9 or newer. A Claude Code mod, installed with claude plugin install textfold@textfold, puts the same five levels on every reply inside the Code tab or the terminal; it needs Claude Code 2.1.287 or later. A Pi extension does the same inside Pi, and a command-line call, python3 -m textfold get, hands an agent one level at a time so the coarse map is read before the full text spends the context window.

The model is one you already have

TextFold looks for a model when it starts and uses what it finds. Claude Code, Codex, Pi and WorkBuddy are detected without a key. A ChatGPT plan is used through a small sign-in bridge that runs locally and needs Node.js. One click authorises OpenRouter and makes its free models available. An API key covers Anthropic, OpenAI, DeepSeek, Qwen, Zhipu GLM, Kimi, MiniMax, Google and Ollama, or any OpenAI-compatible endpoint. A different, stronger model can be set for answering questions about the text, under Settings.

A page can be saved as a single HTML file that opens anywhere without a server, with the levels, chosen notes and the original PDF beside the text. Notes written in the margin stay anchored to the sentences they mark as the reader zooms out. A reading-together feature is not in the project; everything happens on one computer, with the model the reader connects, and the text format the project documents under docs/format.md lets other programs read and write folded text in the same shape.

Source

chenxiachan/textfold on GitHub.