TextFold

Fold any text, from a one-screen map to every word

_{Formerly PaperFold}

English · 简体中文

Paste a link: any web page · arXiv · PubMed Central · bioRxiv · medRxiv · Wikipedia · GitHub
Drop files, one or several: PDF · Word · PowerPoint · Excel · EPUB · HTML · CSV · pictures · Markdown · text

Try the demo → or download the app ↓ · What's new ↓

Read with less in your head

Choose the level of detail you need, and lower your cognitive load.

Also for agent output

Your agent writes more than you have time to read. A Claude Code mod folds every reply the same way, one key per level, and an extension does the same in Pi. More ↓

One fold, three places

• Read. Paste a link or drop a file; it opens as a map of its sections that you zoom into. Get started ↓
• Fold your agent's replies. One key per level in Claude Code and in Pi, with no second model. In Claude Code and Pi ↓
• Plug it into your workflow. python3 -m textfold get hands your agent or script a document one level at a time, and an open format lets any program read and write folded text. For agents ↓

What you get

• Words that travel. As you zoom, every word that stays slides to its new place and the rest fade around it.

• Zoom under your fingers. Stop halfway, turn back, or open one paragraph alone.

• Beside the original. Open a paper's or a PDF's own pages next to the text, each paragraph shown where it stands on its page.

• Ask AI about exactly what you read. Select a passage, or ask about the whole text. The answer comes structured: the short answer first, then the evidence, each point citing the paragraph it rests on, one click away.

• Notes that zoom with the text. Highlight and write in the margin; what you marked stays as you zoom out, and is saved with the page.

• Your language. Read in many languages, translated sentence by sentence, with every formula as typeset.

• Your library, in order. Search it with ⌘K, keep documents in folders (drag one in, or right-click it), and list it by kind, by when each came in, or by when you last opened it.

Highlight and write in the margin…	…and keep what you marked as you zoom out.

Works on your Zotero

Click Zotero library, under the drop zone, and your library is there as Zotero shows it: your collections, your group libraries, a search. Read opens any item.

• An item on arXiv opens from arXiv. Any other opens from its PDF, with the PDF beside the text, or from the EPUB, saved web page or Word file it holds. An item with only a DOI opens from Europe PMC.
• It keeps what you entered in Zotero: the title, the authors and the citekey.
• It stays on your computer. TextFold asks the Zotero app you already run, the way citation plugins do: no account, no key, nothing sent to Zotero's servers, and nothing in your library changed.

To turn it on, in Zotero 7: Settings › Advanced › Allow other applications on this computer to communicate with Zotero.

Get started

Download the app for macOS (Apple silicon or Intel) or Windows. Everything it needs is inside, Python included.

Or install the Python package, on any system and for your own pipeline:

python3 -m pip install textfold
python3 -m textfold serve

Open http://localhost:3017. The first time, connect a model; then paste a link or drop a file, and pick a language. You need Python 3.9 or newer (the one macOS ships with works).

What you read is kept in your system's folder for app data, shared with the desktop app.

git clone https://github.com/chenxiachan/textfold.git
cd textfold
python3 -m pip install -r requirements.txt
python3 -m textfold serve

Connect a model

TextFold uses a model you already have, and there is no account to create. It looks for one when it starts; if it finds none, the first page offers the ways in.

Your questions can go to a stronger model than the one that writes the levels: Settings › Models › Answering questions, with how hard it reasons.

In the reader

For agents and scripts

python3 -m textfold get gives a document one level at a time, coarsest first, so an agent reads the map before it spends its context on the whole text:

python3 -m textfold get https://arxiv.org/abs/2512.23916        # the map: every section's topic
python3 -m textfold get 2512.23916 --level key --section 3      # one section: its points, each with its reason
python3 -m textfold get report.pdf --json                       # the whole document, in the TextFold format

A document not folded yet is folded first, with the model you connected. Every line a model wrote is marked as such and anchored to the author's sentences it rests on: docs/format.md describes the format, for any program that reads or writes folded text.

A running TextFold (the app, or python3 -m textfold serve) says where it is in server.json, in its data folder (~/Library/Application Support/TextFold on a Mac, %LOCALAPPDATA%\TextFold on Windows, ~/.local/share/textfold on Linux, the clone when run from one): {"format": "textfold/server", "version": 1, "url": "http://127.0.0.1:<port>", "port": <port>, "pid": <pid>, "textfold": "<version>", "started": "<UTC time>"}. The file goes when the server stops; one left behind (a server ended outright) names an address that answers nothing.

An agent in a browser that offers WebMCP, a draft web standard (tried in Codex's own browser), finds TextFold's own operations on its pages as tools: it reads the passage you are reading or selected, and, when you ask, saves a note on an exact passage, imports a file or a Zotero item, or sorts your folders. Nothing changes in a browser without it. docs/webmcp.md lists the tools. Contributed by @peitianyi23-sudo.

Save a page

Save, at the top of the reader, keeps a page as one HTML file that opens anywhere without a server: its levels, the notes you choose and, if you like, its original PDF to read beside it. It can also keep the whole text as Markdown at the level you are reading, your notes under their paragraphs (at Topic, an outline that a mind-map plugin draws), or only your notes, as a page, a PDF, Markdown for Obsidian or JSON. From the command line, the page without your notes:

python3 -m textfold export 2512.23916

TextFold is a tool, and its license covers its code. When you share a file, follow the terms of the original text, shown under its title.

In Claude Code and Pi

TextFold puts the five levels on every reply in Claude Code and in Pi. Press 5 for the answer in one sentence, 3 for the point of every paragraph, 1 for every word (in Pi, hold Alt). Nothing is paraphrased, so at every level you read exactly what the model said.

In Claude Code: install it, and see what each level gives you

It needs Claude Code 2.1.287 or later, in the terminal or in the Code tab of the Desktop app.

claude plugin marketplace add chenxiachan/textfold
claude plugin install textfold@textfold

Key	Level	What you get
1	Full	every word Claude wrote
2	Brief	each point, with the reason behind it
3	Points	just the points, at a glance
4	Outline	the map of a long reply: its sections, and what each one says (a reply without headings shows its points)
5	Gist	the answer, in one sentence

Type the digit while the prompt is empty, or pick a level in the row above the prompt. Short replies stay whole: one under about 80 words, or one a fold would barely shorten, always shows as written.

Start folded. A session starts at Full. To start at another level, run /config and set Starting level; in the session, the row above the prompt and the keys 1 to 5 still change it.

Why you can trust a fold. TextFold asks Claude to lead with the point: the answer first in every reply, the key sentence first in every paragraph. A fold keeps what Claude put first. There is no second model and no summary, so a fold costs nothing and takes no time.

More in plugins/textfold.

In Pi: install it, and see it fold

An extension puts the same five levels on every reply in Pi. It needs Pi 1.2.0 or later, on Node 22.15 or later.

git clone https://github.com/chenxiachan/textfold.git
pi install ./textfold/plugins/pi-textfold

Hold Alt and type the digit, click a level in the row above the prompt, or type /fold and a level: /fold 3, /fold points, /fold off. Every reply folds in place, the earlier ones too, and so does the model's thinking, so a long coding turn reads in a few lines. The level stays with the session, and /fold calibrate teaches the keys on any keyboard. On a Mac, let the terminal send Option as Meta, or Option+1 types ¡. It folds with the mod's own engine, so a reply reads the same at each level in Claude Code and in Pi.

More in plugins/pi-textfold. Contributed by @jlamweil.

What's new

• October 22, 2026: figures as TeX drew them, and full screen. In an arXiv paper, a drawing whose words the browser sets itself (a TikZ picture) can have them run out of its boxes, on arXiv's own page too. A button under such a drawing takes the paper's figures of that kind from its PDF, as TeX set them, and another puts them back; nothing is downloaded until you press it. Full screen in Aa fills the screen, and the Topic map opens beside the original pages too. At Brief, the words that go on to Key stand out by a step, not a jump, and the paper style keeps the default's text colours.
• October 22, 2026: lighter on your disk. Pictures are kept as WebP as they come in: an arXiv paper's figures take a fifth of the space or less, and download four at a time, so a new paper opens sooner. Pictures already in your library stay as they are. The app's copy of the pictures and PDFs it showed stays under 32 MB, and a downloaded update is removed once it is installed. A deleted document stays in papers/.trash for 30 days, then it is erased. At Key and Takeaway, a figure of several panels shows them side by side, no longer one under another and cut off.
• October 21, 2026: a reader to stay in. Aa has light or dark (or your system's) and a paper style, quieter than white, for either; and the line spacing. A text opens where you left it. F leaves the text alone on the screen. The sidebar can be pinned beside the text.
• October 21, 2026: tools for an agent in your browser. In a browser that offers WebMCP, an agent finds TextFold's own operations as tools: it reads the passage you selected and, when you ask, saves a note on it, imports a file or a Zotero item, or sorts your folders. The tools. Thanks to @peitianyi23-sudo.
• October 21, 2026: Topic as a mind map. At Topic, Map beside the levels draws the sections and their topics as a mind map: click a topic to read its paragraph, click a section to fold it. Headings stand apart more clearly, Aa can hide the margins (the links, the AI answers, your notes), and Save keeps the whole text as Markdown at the level you read. On Windows, the installer asks where to install.
• October 20, 2026: your type, and several documents at once. Aa at the top sets the text's size, the column's width and serif or sans-serif, for every text. In the sidebar, ⌘-click (Ctrl-click) or Shift-click picks several documents to move into a folder or delete together, and a delete can be undone for a few seconds.
• October 20, 2026: several files at once. Drop or choose several files together: they are read, and their levels written, one after another while you go on reading, and the sidebar shows each.
• October 20, 2026: TextFold and Obsidian. TextFold Bridge, an Obsidian plugin, links your notes to places in TextFold: a paragraph, a section, a note. Drag a paragraph, a section or a selection from the reader into a note and it keeps a link back to where it came from. A note of your own folds at the five levels inside Obsidian, and after an edit it folds again in place, its comments found again in the new text. Thanks to @adalach.
• October 20, 2026: Pi folds the thinking too. In Pi, the model's reasoning folds with its reply, so a long coding turn reads in a few lines. Click a level in the row to fold, keep your level when you resume a session, and teach the keys once with /fold calibrate on any keyboard. Thanks to @jlamweil.
• October 19, 2026: more languages to read in. Swedish, Danish, Norwegian, Finnish, Dutch, Polish, Czech, Ukrainian, Turkish, Indonesian and Vietnamese join the languages a text can be read in. Where the interface has no words for a language, it is in English. Thanks to @Kullenej.
• October 15, 2026: Pi's replies fold too. An extension for Pi puts the five levels on every reply, with the same fold as in Claude Code. Thanks to @jlamweil.
• October 15, 2026: PDFs open without the engine. A PDF with its own text, as most papers have, opens at once, its figures cut from the page; only a scan still needs the PDF engine. An arXiv paper with no HTML version is read from its PDF. Thanks to @sylvainHellin.
• October 15, 2026: your library, and your questions. Search the library, keep documents in folders, and list them by when they came in or were last opened. A question has its own send and stop, any turn of a conversation can be deleted, and questions can go to a stronger model than the one that writes the levels. Thanks to @suhyungJang.
• October 14, 2026: an app for Mac and Windows. Download TextFold for macOS (Apple silicon or Intel, signed and notarized by Apple) or Windows, with everything it needs inside. It finds Claude Code, Codex or Pi on your computer, and updates itself when you say so. Download ↓

Contributors

Contributions are welcome: start with CONTRIBUTING.md.