ipynbtopdf
Live at ipynbtopdf.xyz.
Converts Jupyter .ipynb notebooks to PDF entirely in the browser. No backend, no upload,
no LaTeX toolchain — the notebook is parsed and rendered client-side, then handed to the
browser's own print engine so the PDF keeps real, selectable text.
Running it
npm install
npm run dev
Pushing to master deploys. After a deploy that adds or changes pages, npm run indexnow sends every sitemap URL to IndexNow (Bing and friends) so they recrawl right away; the key file it needs lives in public/.
How the conversion works
src/lib/ipynb.ts is the whole converter:
- Parse — accepts nbformat 4 and upgrades the legacy nbformat 3 layout (cells under
worksheets, input instead of source, pyout instead of execute_result, MIME data
as bare png/latex/html keys).
- Render — Markdown via
markdown-it, code via highlight.js, maths via KaTeX, and
outputs by MIME preference: text/html → text/latex → SVG → raster → plain text. LaTeX
is tried first so a figure that ships both TeX and a bitmap becomes crisp vector maths,
falling back to the bitmap when KaTeX cannot typeset it.
- Sanitize — DOMPurify, with
<use> re-allowed (matplotlib draws every marker with it)
but its references confined to same-document fragments.
Things that look like bugs but are not
- Every inlined SVG gets its ids prefixed. Two exports of one figure carry identical
element ids; once both are on the page the second figure's
url(#…) resolves to the first
one's definitions and it renders blank.
- A lone
\r in stream output collapses the line, \r\n does not. The first is a
progress bar redrawing itself; the second is a Windows line ending.
- Print margins live in
@page, not as padding. Padding applies once across the whole
flow, which leaves content flush against the paper edge at every page break.
@emnapi/core and @emnapi/runtime are devDependencies nothing imports. They are
dependencies of @img/sharp-wasm32, which npm lists in the lockfile but whose own
dependencies it will not resolve while running on Windows. A Linux CI then runs npm ci,
finds them missing, and refuses to install. Declaring them puts them in the lockfile so
the build host is satisfied. Drop them once the lockfile is generated on Linux.
Testing
There is no unit suite. The converter is checked against real notebooks — the corpus at
https://github.com/odewahn/ipynb-examples is a good one, and every file in it exercises the
nbformat 3 path.
