docpull 2.4.0

Pull documentation from the web and convert to clean markdown

docpull

Security-hardened, browser-free crawler that turns static documentation sites into clean, AI-ready Markdown — fast.

docpull uses async HTTP (not Playwright) to fetch server-rendered pages, extracts main content, and writes clean Markdown with source-URL frontmatter — in seconds, with a small install footprint. It won't render JavaScript, but for the large class of docs that don't need it (API references, Python/Go stdlib, most dev-tool docs, OpenAPI specs, Next.js and Docusaurus builds), it is a fast, auditable, sandbox-friendly way to pipe documentation into an LLM context, a RAG index, or an offline archive. SSRF, XXE, DNS-rebinding, and CRLF-injection protections are on by default — a necessity when an AI agent is choosing the URLs.

Install

pip install docpull

# Optional extras
pip install 'docpull[llm]'           # tiktoken for token-accurate chunking
pip install 'docpull[trafilatura]'   # alternative extractor for noisy pages
pip install 'docpull[mcp]'           # run as an MCP server for AI agents
pip install 'docpull[all]'           # everything above

Quick start

# Crawl and save Markdown
docpull https://docs.example.com

# One page, no crawl — the fast path for agents
docpull https://docs.example.com/guide --single

# LLM-ready NDJSON with 4k-token chunks streamed to stdout
docpull https://docs.example.com --profile llm --stream | jq .

# Mirror a site for offline use
docpull https://docs.example.com --profile mirror --cache

Framework-aware extraction

docpull inspects each page before running the generic extractor and can pull content directly from framework data feeds:

Framework	Strategy
Next.js	Parses __NEXT_DATA__ JSON
Mintlify	__NEXT_DATA__ with Mintlify tagging
OpenAPI	Renders openapi.json / swagger.json into Markdown
Docusaurus	Detected and tagged; generic extractor produces Markdown
Sphinx	Detected and tagged; generic extractor produces Markdown

JS-only SPAs with no server-rendered content are detected and skipped with a clear reason (or, with --strict-js-required, reported as an error so agents can route elsewhere).

Agent-friendly features

• --single — fetch a single URL without discovery. Designed for tool loops.
• --stream — NDJSON one-record-per-line, flushed on every page, pipeable.
• --max-tokens-per-file N — split each page into token-bounded chunks on heading boundaries (exact counts with tiktoken, estimate without).
• --emit-chunks — write one file or record per chunk instead of per page.
• --strict-js-required — hard-fail on JS-only pages instead of silently skipping.
• --extractor trafilatura — swap in trafilatura for sites where the default heuristics struggle.

Python API

from docpull import fetch_one

ctx = fetch_one("https://docs.python.org/3/library/asyncio.html")
print(ctx.title, ctx.source_type)
print(ctx.markdown[:500])

Async streaming:

import asyncio
from docpull import Fetcher, DocpullConfig, ProfileName, EventType

async def main():
    cfg = DocpullConfig(
        url="https://docs.example.com",
        profile=ProfileName.LLM,  # chunked NDJSON output
    )
    async with Fetcher(cfg) as fetcher:
        async for event in fetcher.run():
            if event.type == EventType.FETCH_PROGRESS:
                print(f"{event.current}/{event.total}: {event.url}")
        print(f"Done: {fetcher.stats.pages_fetched} pages")

asyncio.run(main())

Single-page from an agent tool:

from docpull import Fetcher, DocpullConfig

async def tool_call(url: str) -> str:
    async with Fetcher(DocpullConfig(url=url)) as f:
        ctx = await f.fetch_one(url, save=False)
        return ctx.markdown or ctx.error or ""

Profiles

docpull https://site.com --profile rag      # Default. Dedup, rich metadata.
docpull https://site.com --profile llm      # NDJSON + chunks + metadata.
docpull https://site.com --profile mirror   # Full archive, polite, cached.
docpull https://site.com --profile quick    # Sampling: 50 pages, depth 2.

MCP server

docpull ships an MCP (Model Context Protocol) server so AI agents can call it directly over stdio:

pip install 'docpull[mcp]'
docpull mcp  # starts the stdio server

Add to Claude Desktop or Claude Code:

{
  "mcpServers": {
    "docpull": {
      "command": "docpull",
      "args": ["mcp"]
    }
  }
}

Tools exposed:

• fetch_url(url, max_tokens?) — one-shot fetch, no crawl
• ensure_docs(source, force?) — fetch a named library (cached 7 days)
• list_sources(category?) — show available aliases (react, nextjs, fastapi, …)
• list_indexed() — what has been fetched locally
• grep_docs(pattern, library?) — regex search across fetched Markdown

User-defined sources live in ~/.config/docpull-mcp/sources.yaml:

sources:
  mydocs:
    url: https://docs.example.com
    description: My internal docs
    category: internal
    maxPages: 200

Output

Markdown files with YAML frontmatter:

---
title: "Getting Started"
source: https://docs.example.com/guide
source_type: "nextjs"
---

# Getting Started
…

NDJSON (one record per page or chunk):

{"url": "...", "title": "...", "content": "...", "hash": "...", "token_count": 842, "chunk_index": 0}

Security

• HTTPS-only, mandatory robots.txt compliance
• SSRF protection: blocks private/internal network IPs, DNS rebinding via connect-time address pinning
• XXE protection via defusedxml on sitemaps
• Path traversal and CRLF header injection guards
• Auth headers stripped on cross-origin redirects

When running with --proxy, DNS pinning is delegated to the proxy. Pass --require-pinned-dns to refuse this configuration and keep the connector- level SSRF guarantees in effect.

Options

Run docpull --help for the full list. Highlights:

Core:
  --profile {rag,mirror,quick,llm,custom}
  --single                Fetch one URL (no crawl)
  --format {markdown,json,ndjson,sqlite}
  --stream                Stream NDJSON to stdout

LLM / chunking:
  --max-tokens-per-file N
  --tokenizer NAME        tiktoken encoding (default cl100k_base)
  --emit-chunks           One file/record per chunk

Content extraction:
  --extractor {default,trafilatura}
  --no-special-cases      Disable framework extractors
  --strict-js-required    Error on JS-only pages

Cache:
  --cache                 Enable incremental updates
  --cache-dir DIR
  --cache-ttl DAYS

Performance

End-to-end numbers from tests/benchmarks/test_10k_pages.py against a synthetic 10,000-page localhost site (RAG profile, max_concurrent=50, HTTP keep-alive, 5% injected duplicate content):

Metric	Value
Total wall time	~27 s
Discovery (sitemap parse)	~80 ms
Fetch + convert + save	~27 s
Per-page latency p50 / p95 / p99	~2.6 / 4.6 / 5.3 ms
Peak RSS delta from baseline	~28 MB
Cache manifest size on disk	~3.4 MB
Duplicates detected (5% injected)	499 / 500

Reproduce with make benchmark (requires aiohttp; runs the gated benchmark in tests/benchmarks/ and prints a JSON line you can pipe into trend tooling).

Troubleshooting

docpull --doctor              # Check installation
docpull URL --verbose         # Verbose output
docpull URL --dry-run         # Test without downloading
docpull URL --preview-urls    # List URLs without fetching

Links

• Website
• PyPI
• GitHub
• Changelog

License

MIT