WDF · Web Document Format

Open document format · Core 0.1 draft

Web Document Format (.wdf) What you read is what AI reads.

WDF is an open, web-native document format built for accessibility: readable on any screen, legible to machines (AI bots), verifiable by construction. One file carries the text, the data and the original document — and proves they agree.

  • Built for accessibility — made for public administration
  • ZIP + HTML + JSON, nothing invented
  • SHA-256 integrity, verified offline
  • Spec CC-BY 4.0 · Code Apache-2.0

Why?

A document has two audiences now. People and Agents. Every format we publish in was designed for exactly one of them.

The reader changed. The format didn't.

PDF freezes the print view and throws away the structure the document had while it was written. A web page keeps the structure, but it is not a file: it changes under your citation and dies with its server.

You can't parse your way to proof.

An entire industry re-guesses structure from PDFs before feeding them to language models. A parser gives you an interpretation. No pipeline can prove that what the machine read is what you saw.

Keep the PDF. Gain the rest.

WDF does not fight the old world. The author's own PDF and the original Word file travel inside the package, next to the real structure — so nothing you deliver today is lost.

Built for accessibility.

Public bodies must publish documents everyone can read — on a phone, with a screen reader, at any zoom (EU Directive 2016/2102, European Accessibility Act). A WDF is semantic HTML on a closed whitelist: headings, lists, real tables, text that reflows, alt text required. The structure is there from the start, nothing to tag afterwards; what remains — descriptions, contrast, plain language — stays the author's job, and stays checkable.

Read why we're doing this →

Live

The Reader, with a real document in it.

Toggle Human / Agent, click any paragraph, copy a citation: the same element lights up in both views. The badge means the file verified — here, on your machine, with no server involved.

A fictional municipal decree, written for demonstration. Open full screen →

What's in the file

One file, every reader.

  • One file

    Text, typed data and the original document travel together in a single portable .wdf.

  • Verified

    SHA-256 hashes and a byte-deterministic derivation. Verify offline, with no authority to call.

  • AI & MCP

    Agents connect to the structure through the MCP server. They never scrape an image of it.

  • Any screen

    Real text that reflows on a phone, a tablet, a desktop — and reads aloud correctly.

  • Citations that resolve

    A wdf: URI opens the exact paragraph it came from. Not “page 2”.

  • Typed data

    Tables are bound to typed datasets: numbers you can compute on, not pictures of digits.

  • .wdf + PDF view

    The author's own PDF travels inside and shows as the Original view, identical to paper.

  • .wdf + original download

    The Word file is inside too, and comes out unchanged.

  • Save the page

    The browser extension turns any web page into a verifiable file you can send. Get it for Chrome →

  • Works offline

    A document never touches the network. Not to open, not to verify. Ever.

  • Accessible

    Semantic HTML on a closed whitelist: the structure accessibility needs, built in rather than tagged on afterwards.

  • Open format

    Spec CC-BY 4.0, reference implementation Apache-2.0. Nothing to license, nobody to ask.

Perfect for

Where the old formats hurt most.

Public sector

Public documents everyone can read

Decrees, reports, calls: readable on a phone, verifiable by any citizen, with the PDF still inside for the archive. The structure accessibility rules demand, built in.

See the decree →

Web

Save a page, send a file

One click in the browser turns the article you are reading into a document that opens anywhere and declares where it came from. Link rot ends here.

Get the extension →

AI

Knowledge bases that cite

A dossier converted to WDF gives an agent stable identifiers and verifiable citations through MCP — no parsing, no hallucinated page numbers.

Read the MCP session →

Office

Publish a Word document, the new way

report.docx in, report.wdf out: structure, typed tables, the original file and the author's PDF, all in one package.

Drop a .docx into the Reader →

Get started

Three ways in.

Read

WDF Reader

An installable app: drop a .wdf, a .docx or an HTML export, read it, verify it, cite it. Works offline once installed.

Click Install in the Reader (Chrome, Edge). Then .wdf files open with a double click.

Save

Browser extension

Save as WDF captures the page you are on as a .wdf or a standalone .wdf.html with the reader built in. No servers, no data collected.

Firefox: in review. Privacy policy.

Build

Tools on npm

Validate, convert and extract from the command line, in CI, or from your own code. Same input, byte-identical package on any platform.

# validate any document
npx @wdf-dev/cli validate document.wdf
# Word → WDF, keeping the original and its PDF
npx @wdf-dev/cli import report.docx \
    --with-source --with-pdf report.pdf
# give an agent the document
claude mcp add wdf -- npx -y @wdf-dev/mcp

Source on GitHub · Spec, ~20 pages · PDF vs WDF, one prompt