getdocs

Turn any documentation site into a clean, local markdown copy your coding agent can actually read.

getdocs crawl https://example.com/docs -o ./out

Coding agents are only as good as the docs they can see. Pointing an agent at a live docs URL means it burns tokens on nav bars, cookie banners, and HTML chrome — or can't reach the page at all. getdocs gives the agent a local, offline, markdown mirror instead: the actual content, structured to match the original site, ready to drop into a repo or feed to a model.

Why getdocs

• Richer context for coding agents. A local copy is greppable, indexable, and always available — the agent reads the whole library at once instead of fetching one rendered page at a time. No rate limits, no network flakiness, no JS that won't hydrate.
• Clean markdown → fewer tokens. Each page is reduced to its content (the nav, headers, footers, and ad chrome stripped) and written as plain markdown. Agents consume it directly, and you spend tokens on docs, not <div> soup.
• Structure preserved. Files mirror the URL hierarchy (example.com/docs/auth → out/docs/auth.md), each with YAML frontmatter (url, title, crawled_at, status), plus a crawl.json Manifest that captures the site's nav order and reading order — so an agent can follow the docs in the order the authors intended.
• Source-first: clone over crawl. If the docs site is open-source, getdocs detects the "Edit this page" link, clones the repo, and serves you the original markdown source instead of scraping HTML — the highest-fidelity copy there is. Falls back to crawling automatically when there's no repo.

When to reach for it

• Coding against an unfamiliar library or API. Mirror its docs into your repo (or a scratch dir) so your agent can ground its answers in the real reference instead of hallucinating from memory.
• RAG / knowledge bases. Get a clean markdown corpus to chunk and embed, without writing a bespoke scraper-and-cleaner for every site.
• Offline or air-gapped work. Take a docs set with you; read and search it with no network.
• Pinning a version. Snapshot today's docs so your agent isn't tripped up when the upstream site changes underneath you.
• Migrating or archiving docs. Pull an entire site down as markdown to move, diff, or keep.

Output

out/
├── crawl.json            ← the Manifest: nav order, reading order, what ran
└── docs/
    ├── index.md
    ├── auth.md
    └── guide/
        └── intro.md

---
url: https://example.com/docs/auth
title: Authentication
status: 200
crawled_at: 2026-06-12T10:00:00Z
---

# Authentication
...

Sitemap discovery, JavaScript rendering, source-repo cloning, polite throttling, JSONL output, and resumable crawls are all built in — see docs/USAGE.md.

Use with your agent

getdocs is built to be driven by a coding agent: it's an ordinary CLI whose out/ tree + crawl.json Manifest is the return value (no MCP server, no job protocol — see ADR-0007). Two patterns cover most uses.

Synchronous — fetch one docs section. Scope defaults to the seed's host + path prefix, so pointing at a subtree fetches just that subtree. This blocks until done and works under any agent:

getdocs crawl https://example.com/docs/auth -o ./out --summary-json

Background — mirror a whole site. Kick the crawl off as a background task and keep working. Under Claude Code the agent is resumed automatically when the crawl finishes; every other agent surveyed blocks or polls the output path instead (this is a harness feature, not a getdocs one):

getdocs crawl https://example.com/docs -o ./out --summary-json &

Read the summary, branch on the Outcome. Every run ends with a one-line summary on stderr; --summary-json adds a machine-readable object discriminated by outcome. A run produces exactly one Outcome — a Crawl or a Clone:

// outcome: "crawled" — getdocs scraped the rendered site
{ "outcome": "crawled", "status": "ok", "pages": 42,
  "output_dir": "./out", "manifest": "./out/crawl.json", "truncated": false }

// outcome: "cloned" — the docs were open-source, so getdocs cloned the repo
// (no pages / no manifest: a Clone is not a Crawl)
{ "outcome": "cloned", "status": "ok", "repo": "acme/docs",
  "output_dir": "./out/docs", "mkdocs_config": "./out/mkdocs.yml" }

Have the agent switch on outcome:

• crawled → grep and read the Pages under output_dir and follow the nav / reading order in manifest (crawl.json).
• cloned → you have the original markdown source; serve it with mkdocs serve -f <mkdocs_config> (or just read the files under output_dir).
• status: "truncated" → the crawl hit its page cap; re-run with a higher --limit (or 0 for unlimited) to get the rest.

Install

Requires Python 3.12+.

pip install getdocs

Or from source, for the latest unreleased changes:

git clone https://github.com/jonbakerfish/getdocs
cd getdocs
pip install -e .

That's enough to crawl. Two optional pieces unlock more:

# JavaScript rendering — the headless browser used to hydrate SPA docs
playwright install chromium

# Serve a crawled/cloned copy locally as a browsable site
pip install mkdocs mkdocs-material

git must be on your PATH for source-first cloning (it almost always already is); without it, getdocs simply falls back to crawling. To run the optional API service, install the server extra: pip install "getdocs[server]".

Development

git clone https://github.com/jonbakerfish/getdocs
cd getdocs
pip install -e ".[dev]"
pytest

Responsible use

getdocs is a tool; how you point it is on you. By default it honors robots.txt, throttles itself politely, and identifies itself honestly in the User-Agent (getdocs/<version> (+project-url)) — please keep it that way. For high-volume crawls, add --contact you@example.com so site operators can reach you (it's appended to the User-Agent; optional but courteous).

getdocs is intended for personal, reference, and agent/RAG use on documentation you have the right to access. You are solely responsible for complying with each site's Terms of Service, its robots.txt, applicable law, and the copyright of the content you fetch — getdocs is provided as-is, with no warranty (see LICENSE). Crawled documentation belongs to its authors: use it for your own reference, but don't redistribute someone else's docs as your own. Crawl only what you have the right to.