Documents → Markdown (markitdown in the browser)

A single-page web app that converts a wide range of documents to Markdown entirely in your browser — no server, no uploads, your files never leave the device.

It runs Microsoft markitdown (Python) inside Pyodide — CPython compiled to WebAssembly — and lets markitdown's own built-in converters do the work.

Live site: https://spitzerd.github.io/doc-to-markdown/

Supported formats

Everything markitdown can handle fully offline inside the WASM sandbox:

Category	Extensions	Backend
PDF	.pdf	pdfplumber / pdfminer.six
Word	.docx	mammoth
PowerPoint	.pptx	python-pptx (+ lxml)
Excel	.xlsx	openpyxl (+ pandas)
Excel (legacy)	.xls	xlrd (+ pandas)
Outlook	.msg	olefile
EPub	.epub	zipfile + defusedxml + HTML
HTML	.html, .htm	BeautifulSoup / markdownify
Tabular / data	.csv, .json, .ipynb	built-in converters
Markup / text	.xml, .txt, .md, …	plain-text converter
Archives	.zip	iterates and converts each entry

Not supported in the browser (they need native binaries or network calls that don't exist in the sandbox): image OCR / EXIF (needs the exiftool binary or an LLM key), audio/video transcription, and URL sources such as YouTube, RSS and Wikipedia.

How it works

Piece	Role
index.html	UI: drag-and-drop, status, result with copy/download
app.js	Talks to the worker, manages UI state and the accepted-extension list
worker.js	Loads Pyodide, installs markitdown wheels, runs conversions

On first visit the browser downloads the Pyodide runtime (cached by the browser/CDN for a year), then:

1. pyodide.loadPackage([...]) — native WASM builds and pure-Python packages that ship in Pyodide's own lockfile: pillow, cryptography, lxml, pandas, numpy, xlrd, beautifulsoup4, charset-normalizer, …
2. micropip.install([...], deps=False) — markitdown 0.1.7 plus the pure-Python wheels Pyodide does not ship (defusedxml, markdownify, openpyxl, python-pptx, mammoth, olefile, …) straight from PyPI.
3. A stub magika module is injected before importing markitdown: magika needs onnxruntime, for which no WASM build exists. The stub always reports "unknown", so markitdown falls back to extension / MIME based format detection — we pass the real extension of the dropped file.

The dropped file is read as bytes and handed to MarkItDown.convert_stream(..., stream_info=StreamInfo(filename=..., extension=...)), so the exact same routing markitdown uses on the desktop is exercised in the browser.

Deliberate version pins

• charset-normalizer is loaded from the Pyodide lockfile (3.4.x). Do not install 3.5.x — it ships a Rust pyemscripten wheel that hard-aborts when imported under Pyodide.
• pdfplumber==0.11.6 + pdfminer.six==20250327 — the last pdfplumber series compatible with a pure-Python install (0.11.10+ pulls native pypdfium2).
• markitdown==0.1.7 with deps=False — so magika (and its onnxruntime requirement) is never fetched; the remaining graph is installed by hand.

Run locally

python -m http.server 8080
# open http://localhost:8080

Any static server works; there is no build step. (Pyodide needs SharedArrayBuffer -free loading here, so a plain static server with the default COOP/COEP-less setup is sufficient.)

Deploy to GitHub Pages

.github/workflows/deploy.yml publishes the repository root as a static Pages site (Jekyll disabled via .nojekyll) on every push to main.

Notes & limits

• First run downloads the runtime plus the Office/Excel/Word wheels; afterwards everything is cached and conversion works offline too.
• Conversion speed is comparable to desktop markitdown; a typical PDF converts in seconds (a 6.8 MB arXiv paper takes ~17 s in Chrome on desktop).
• Output quality matches desktop markitdown for the supported formats. Scanned / image-only PDFs have no extractable text, and the same is true of the CLI.