docmark

Convert hosted documentation sites to local Markdown — built for feeding LLMs and AI Skills.

Currently optimized for Mintlify-hosted docs (Anthropic, Polymarket, many web3 / crypto sites), which expose the source markdown of any page at <url>.md. The architecture is built around a single downloader strategy, so other doc platforms (Docusaurus, MkDocs, GitBook, ReadMe, generic HTML) can be added without rewriting the rest of the pipeline.

Install

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e .

Use

docmark https://docs.polymarket.com/sitemap.xml --output ./output/polymarket

Or without installing:

python -m docmark https://docs.polymarket.com/sitemap.xml --output ./output/polymarket

Options

Flag	Default	Description
--output, -o	output	Directory to write markdown files into
--concurrency, -c	10	Parallel downloads
--include-locales	off	Include localized variants (/cn/, /es/, ...). Filtered out by default.
--include	none	Only crawl URLs whose path starts with this prefix
--exclude	none	Skip URLs whose path starts with this prefix (repeatable)
--timeout	30	Per-request timeout in seconds

Examples

Only API reference pages, higher concurrency:

docmark https://docs.polymarket.com/sitemap.xml -o ./out -c 20 --include /api-reference/

Include Chinese variant and skip the /builders/ section:

docmark https://docs.polymarket.com/sitemap.xml -o ./out --include-locales --exclude /builders/

How URL paths map to files

https://docs.polymarket.com/                           -> output/index.md
https://docs.polymarket.com/quickstart                 -> output/quickstart.md
https://docs.polymarket.com/api-reference/trade/cancel-all-orders
                                                       -> output/api-reference/trade/cancel-all-orders.md

How it works

Mintlify renders HTML for users, but also serves the raw MDX source whenever a request appends .md to a page URL:

https://docs.example.com/quickstart       -> rendered HTML
https://docs.example.com/quickstart.md    -> raw markdown source

The crawler reads the site's sitemap.xml, requests <url>.md for every entry in parallel, and writes each response to disk preserving the URL path. No HTML parsing, no headless browser, no conversion loss — output matches what the docs author wrote.

Detecting Mintlify

A site is likely Mintlify if any of these hold:

• <meta name="generator" content="Mintlify"> in the HTML
• Assets served from mintcdn.com
• A llms.txt or llms-full.txt file exists at the site root
• Appending .md to a doc URL returns plain markdown (not HTML)

If .md requests return HTML, the site is not Mintlify and a different strategy is needed.

Supported platforms

Platform	Status	Strategy
Mintlify	Implemented	Append .md to each page URL
Docusaurus	Possible	Fetch source .md / .mdx from the docs repo on GitHub
MkDocs	Possible	Same — fetch source from the GitHub repo
GitBook	Possible	GitBook API (with token), or HTML scrape
ReadMe	Possible	ReadMe API (with token), or HTML scrape
Generic / custom	Possible	HTML scrape (markdownify or html2text)

The downloader (src/docmark/downloader.py) is the only piece that knows about a specific platform. Adding a new strategy means writing a small module with a fetch(page_url, client) -> DownloadResult function and wiring it as a --strategy choice in the CLI. Sitemap parsing, filters, file writing, and concurrency stay untouched.

Strategies are added on demand — when a concrete site needs them — not speculatively.

Notes

• Sitemap-driven. URLs not listed in sitemap.xml are not crawled.
• Pages are saved as raw MDX. Mintlify components (<Steps>, <Tabs>, <CardGroup>, ...) are preserved verbatim — Claude and other LLMs read them fine.
• A best-effort fetch of llms.txt and llms-full.txt from the site root is included.

License

MIT — see LICENSE.