Skip to content

AIMLPM/markcrawl

MarkCrawl by iD8 🕷️📝

Turn any webpage or website into clean Markdown for LLM pipelines — in one command.

CI PyPI Version License markcrawl MCP server

Latest: v0.11.1 (2026-05-12) — default aggregator URL filter. See What's New below.

pip install markcrawl
markcrawl --base https://docs.example.com --out ./output --show-progress

MarkCrawl is a crawl-and-structure engine. It fetches one page or crawls an entire website, strips navigation/scripts/boilerplate, and writes clean Markdown files with a structured JSONL index. Every page includes a citation with the access date. No API keys needed.

Everything else — LLM extraction, Supabase upload, MCP server, LangChain tools — is optional and installed separately.

Want a hosted API instead of running locally? Join the waitlist — we're gauging interest.

LLM agents: Load docs/LLM_PROMPT.md as a system prompt to generate correct MarkCrawl commands automatically.

What's New

Install or upgrade with pip:

pip install --upgrade markcrawl
pip show markcrawl | grep Version       # confirm the installed version
markcrawl --help | head -1              # confirm the binary on $PATH is the upgraded one

If markcrawl --help is missing flags you expect (e.g. --screenshot, --seed-file, --smart-sample, --download-images), your local install is stale. Run pip install --upgrade markcrawl against the same Python that owns the markcrawl binary on your PATHhead -1 $(which markcrawl) shows the right interpreter. PyPI is always the source of truth; see CHANGELOG.md for the full release history.

v0.11 highlights (changelog):

  • Aggregator URL filter (default, v0.11.1) — rejects mdBook /print.html and Hugo /_print/ pages during crawl-time URL filtering. These bundle the entire docs tree on a single URL and otherwise dominate retrieval rankings on cosine similarity (markcrawl was returning them in 49% of rust-book and 39% of kubernetes-docs top-5 retrieval slots before the fix; competitors return 0%). Opt out via include_aggregator_pages=True / --include-aggregators.
  • Binary downloads (v0.11.0) — new download_types=["pdf", "docx"] kwarg streams referenced files to <out_dir>/downloads/ with size + content-type guards. Pre-fetch download_filter callback receives URL + anchor text + parent-page context; reject candidates before any HTTP bytes transfer.
  • Local embedder is the default since v0.10.1 — pip install markcrawl ships the full ML stack (torch + transformers + sentence-transformers). Zero API key required for embedding. Override with MARKCRAWL_EMBEDDER=text-embedding-3-small or the embedding_model kwarg if you want OpenAI back.
  • Tenacity-backed HTTP retry — full-jitter exponential backoff (2 s → 30 s, 5 attempts) that honors the server's Retry-After header on 429s.

Where markcrawl stands on the public benchmark, honestly. The independent llm-crawler-benchmarks v1.4 leaderboard measures 7 web crawlers on how well their output supports RAG. Markcrawl ranks 1st on cost ($4,505/yr at 100,000-page scale) but 7th of 7 on answer quality (3.77/5) and retrieval accuracy (MRR 0.341 vs leaders at 0.76). We're actively working to close that gap on three fronts:

  1. v0.11.1 (just shipped) filters out /print.html and /_print/ "whole-book-on-one-page" URLs that were stealing 39–49% of markcrawl's top-5 retrieval slots on documentation sites. Competitors already filter these. Expected MRR improvement: +0.02 to +0.04 on docs-heavy sites (formal measurement pending the next benchmark cycle).
  2. Upcoming releases improve how markcrawl chooses which pages to crawl within its budget — markcrawl's deliberately-narrower crawl strategy (which keeps cost low and signal-to-noise high) is also the main cause of the retrieval gap.
  3. The benchmark itself is being improved — v1.4's test questions were sampled from one specific crawler's output, which structurally penalizes any crawler whose discovery strategy differs from that anchor. The benchmark is being updated so each site's test questions come from the site's own sitemap, independent of any crawler. We expect this fix alone to surface ~5–10% of markcrawl's current "misses" as actually correct answers at different URLs — work shown in our audit notes.

Goal for the next benchmark cycle: move from 7th to mid-pack on retrieval (+0.10 to +0.20 MRR) and answer quality, while keeping the cost-efficiency lead. Honest, measured progress — we publish the numbers either way.

Quickstart (2 minutes)

pip install markcrawl
markcrawl --base https://quotes.toscrape.com --out ./demo --max-pages 5 --show-progress

Your ./demo folder now contains:

demo/
├── index__a4f3b2c1d0.md    ← clean Markdown of the page
├── page-2__b7e2d1f0a3.md
├── ...
└── pages.jsonl              ← structured index (one JSON line per page)

Each line in pages.jsonl:

{
  "url": "https://quotes.toscrape.com/",
  "title": "Quotes to Scrape",
  "crawled_at": "2026-04-04T12:30:00Z",
  "citation": "Quotes to Scrape. quotes.toscrape.com. Available at: https://quotes.toscrape.com/ [Accessed April 04, 2026].",
  "tool": "markcrawl",
  "text": "# Quotes to Scrape\n\n> "The world as we have created it is a process of our thinking..." — Albert Einstein\n\nTags: change, deep-thoughts, thinking, world..."
}

Schema — every page in pages.jsonl has these fields:

Field Type Description
url string Original URL fetched.
title string Page title from <title> (or first H1 if missing).
crawled_at string (ISO 8601) UTC timestamp of when the page was fetched.
citation string Pre-formatted academic-style citation including access date.
tool string Always "markcrawl". Helps when merging output from multiple crawlers.
text string Clean Markdown content (nav/footer/scripts stripped).
downloads array (optional) Present when download_types is set; one entry per saved binary: {url, path, size_bytes, content_type}.
images array (optional) Present when --download-images is set; lists saved image paths.
screenshot string (optional) Present when --screenshot is set; relative path to the PNG/JPEG capture.

Common Recipes

Runnable examples for the most common patterns:

  • Single-page scrapes — including JS-rendered pages (React, Vue, YouTube)
  • Whole-site crawls — docs, blogs, subsections; resume interrupted runs
  • URL filtering--exclude-path, --include-path, --dry-run, smart sampling
  • Extraction backends — BS4 (default), trafilatura, ensemble, ReaderLM-v2
  • Binary downloads — images, PDFs (with pre-fetch filter callbacks), DOCX
  • Screenshots — full-page or cropped, PNG or JPEG
  • End-to-end use cases — competitive analysis, RAG chatbot, API-docs → code-gen

Full recipes with copy-paste commands and expected outputs: docs/RECIPES.md.

How it compares to other crawlers — decision matrix

Pick this tool when…

If you need… Use… Why
Clean Markdown for LLM/RAG ingestion, run locally, no API keys MarkCrawl Default install bundles local embedder ($0 API spend); strips nav/scripts; produces JSONL with citations out of the box
A hosted scraping API (no infra to run) FireCrawl SaaS option; pay-per-call; outsources crawling entirely
AI-native crawling with built-in LLM extraction Crawl4AI Deeper LLM-extraction primitives; built-in Playwright
Massive distributed crawling (millions of pages, custom pipelines) Scrapy Battle-tested framework; rich plugin ecosystem; spider architecture
JavaScript-heavy automation without framework overhead Playwright (direct) Lower-level control over browser automation
Sites behind login/auth or aggressive bot protection None of the above (build custom) See When NOT to use MarkCrawl; same constraints apply to most public crawlers

Feature comparison

MarkCrawl FireCrawl Crawl4AI Scrapy
License MIT AGPL-3.0 Apache-2.0 BSD-3
Install pip install markcrawl SaaS or self-host pip + Playwright pip + framework
Output Markdown + JSONL Markdown + JSON Markdown Custom pipelines
JS rendering Optional (--render-js) Built-in Built-in Plugin
LLM extraction Optional add-on Via API Built-in None
Local-only operation ❌ (SaaS)
Citations + timestamps in output Partial Manual
Best for Single-site crawl → clean Markdown Hosted scraping API AI-native crawling Large-scale distributed

MarkCrawl's niche is focused-scope RAG ingestion — narrow crawls of docs/blogs/product sites that produce LLM-ready Markdown with minimal junk. For broader scope or bigger scale, the other tools above are stronger choices.

Benchmark results (6 tools, May 2026)

Speed: scrapy+md is fastest (5.0 pages/sec), markcrawl at 2.7. Playwright-based tools average 1.4-2.1 pages/sec.

Output cleanliness: markcrawl has the lowest nav pollution (53 words vs 500+ for others) — less junk in your embeddings.

RAG answer quality: markcrawl scores 3.77/5 on answer quality with the fewest chunks (27,193 total, 2.2x fewer than the most), keeping embedding costs low.

Tool Chunks/page Answer Quality (/5) Annual cost (100K pages, 1K queries/day)
markcrawl 18.7 3.77 $4,505
scrapy+md 31.7 3.68 $5,464
crawl4ai 16.8 4.72 $6,960
colly+md 40.6 4.36 $7,213
playwright 39.0 4.48 $7,320
crawlee 40.5 4.68 $7,467

Full benchmark data: docs/BENCHMARKS.md | Methodology: llm-crawler-benchmarks

Methodology caveat (numbers as of bench v1.4, 2026-05-11): the v1.4 leaderboard sourced test queries from a single high-coverage crawler's output. The bench is actively being updated in v1.5 to source queries from each site's own sitemap independent of any crawler (release notes). Numbers above are single-trial; multi-trial measurement is on the v1.5.1 roadmap. Treat individual rankings as point-in-time signal, not steady-state.

Installation

pip install markcrawl                # Core crawler + chunker + local embedder
                                     # (no API keys required for embedding)

Optional add-ons (tasks beyond the crawl-and-embed core):

pip install markcrawl[js]            # + JavaScript rendering (Playwright)
pip install markcrawl[extract]       # + LLM extraction (OpenAI, Claude, Gemini, Grok)
pip install markcrawl[upload]        # + Supabase upload integration
pip install markcrawl[mcp]           # + MCP server for AI agents
pip install markcrawl[langchain]     # + LangChain tool wrappers
pip install markcrawl[all]           # Everything

For Playwright, also run playwright install chromium after installing.

Lean install (skip the local-embedder dep stack — you'll need an OPENAI_API_KEY and pass embedding_model="text-embedding-3-small" for any embedding work):

pip install --no-deps markcrawl beautifulsoup4 lxml markdownify requests certifi tenacity
Install from source (for development)
git clone https://github.com/AIMLPM/markcrawl.git
cd markcrawl
python -m venv .venv
source .venv/bin/activate
pip install -e ".[all]"

Crawling

markcrawl --base https://www.example.com --out ./output --show-progress

Add flags as needed:

markcrawl \
  --base https://www.example.com \
  --out ./output \
  --include-subdomains \        # crawl sub.example.com too
  --render-js \                 # render JavaScript (React, Vue, etc.)
  --concurrency 5 \             # fetch 5 pages in parallel
  --proxy http://proxy:8080 \   # route through a proxy
  --max-pages 200 \             # stop after 200 pages
  --format markdown \           # or "text" for plain text
  --show-progress

Resume an interrupted crawl:

markcrawl --base https://www.example.com --out ./output --resume --show-progress

Output

Each page becomes a .md file with a citation header:

# Getting Started

> URL: https://docs.example.com/getting-started
> Crawled: April 04, 2026
> Citation: Getting Started. docs.example.com. Available at: https://docs.example.com/getting-started [Accessed April 04, 2026].

Welcome to the platform. This guide walks you through installation...

Navigation, footer, cookie banners, and scripts are stripped. Only the main content remains.

All crawler CLI arguments
Argument Description
--base Base site URL to crawl
--out Output directory
--format markdown or text (default: markdown)
--show-progress Print progress and crawl events
--render-js Render JavaScript with Playwright before extracting
--concurrency Pages to fetch in parallel (default: 1)
--proxy HTTP/HTTPS proxy URL
--resume Resume from saved state
--include-subdomains Include subdomains under the base domain
--max-pages Max pages to save; 0 = unlimited (default: 500)
--delay Minimum delay between requests in seconds (default: 0, adaptive throttle adjusts automatically)
--timeout Per-request timeout in seconds (default: 15)
--min-words Skip pages with fewer words (default: 20)
--user-agent Override the default user agent
--use-sitemap / --no-sitemap Enable/disable sitemap discovery. Use --no-sitemap when you want to scrape a specific page or subsection — without it, large sites (YouTube, GitHub) may discover thousands of unrelated pages via their sitemap
--exclude-path Glob pattern to exclude URL paths (e.g. '/job/*'). Can be repeated
--include-path Glob pattern to include URL paths (e.g. '/blog/*'). Only matching paths are crawled. Can be repeated
--dry-run Discover URLs (via sitemap/links) and print them without fetching content
--smart-sample Auto-detect templated URL patterns and sample from large clusters instead of crawling every page
--sample-size Pages to sample per templated cluster (default: 5, used with --smart-sample)
--sample-threshold Clusters larger than this are sampled (default: 20, used with --smart-sample)
--auto-resume Automatically resume if saved state exists, otherwise start fresh
--cross-dedup Skip pages already seen in previous crawls to the same output directory
--prioritize-links Score discovered links by predicted content yield — crawl high-value pages first
--extractor Content extraction backend: default, trafilatura, ensemble, or readerlm
--download-images Download images from the content area to assets/ and use local paths in Markdown
--min-image-size Minimum image file size in bytes to keep (default: 5000). Smaller images are skipped
--i18n-filter Skip URLs under locale path segments (/fr/, /de-DE/, /zh-Hans/, ...) — generic, no per-domain config
--title-at-top Prepend # {title} to the text field of every JSONL row when not already present — top-MRR RAG recipe

Optional: structured extraction

If you need structured data (not just text), the extraction add-on uses an LLM to pull specific fields from each page.

pip install markcrawl[extract]

markcrawl-extract \
  --jsonl ./output/pages.jsonl \
  --fields company_name pricing features \
  --show-progress

Auto-discover fields across multiple crawled sites:

markcrawl-extract \
  --jsonl ./comp1/pages.jsonl ./comp2/pages.jsonl ./comp3/pages.jsonl \
  --auto-fields \
  --context "competitor pricing analysis" \
  --show-progress

Supports OpenAI, Anthropic (Claude), Google Gemini, and xAI (Grok) via --provider.

Extraction details

Provider and model selection

markcrawl-extract --jsonl ... --fields pricing --provider openai         # default
markcrawl-extract --jsonl ... --fields pricing --provider anthropic      # Claude
markcrawl-extract --jsonl ... --fields pricing --provider gemini         # Gemini
markcrawl-extract --jsonl ... --fields pricing --provider grok           # Grok
markcrawl-extract --jsonl ... --fields pricing --model gpt-4o           # override model
Provider API key env var Default model
OpenAI OPENAI_API_KEY gpt-4o-mini
Anthropic ANTHROPIC_API_KEY claude-sonnet-4-20250514
Google Gemini GEMINI_API_KEY gemini-2.0-flash
xAI (Grok) XAI_API_KEY grok-3-mini-fast

All extraction CLI arguments

Argument Description
--jsonl Path(s) to pages.jsonl — pass multiple for cross-site analysis
--fields Field names to extract (space-separated)
--auto-fields Auto-discover fields by sampling pages
--context Describe your goal for auto-discovery
--sample-size Pages to sample for auto-discovery (default: 3)
--provider openai, anthropic, gemini, or grok
--model Override the default model
--output Output path (default: extracted.jsonl)
--delay Delay between LLM calls in seconds (default: 0.25)
--show-progress Print progress

Output format

Extracted rows include LLM attribution:

{
  "url": "https://competitor.com/pricing",
  "citation": "Pricing. competitor.com. Available at: ... [Accessed April 04, 2026].",
  "pricing_tiers": "Starter ($29/mo), Pro ($99/mo), Enterprise (contact sales)",
  "extracted_by": "gpt-4o-mini (openai)",
  "extraction_note": "Field values were extracted by an LLM and may be interpreted, not verbatim."
}

Optional: Supabase vector search (RAG)

Chunk pages, generate embeddings, and upload to Supabase with pgvector:

pip install markcrawl[upload]

markcrawl --base https://docs.example.com --out ./output --show-progress
markcrawl-upload --jsonl ./output/pages.jsonl --show-progress

Requires SUPABASE_URL, SUPABASE_KEY, and OPENAI_API_KEY. See docs/SUPABASE.md for table setup, query examples, and recommendations.

Optional: agent integrations

MarkCrawl includes integrations for AI agents. Each is an optional add-on.

MCP Server (Claude Desktop, Cursor, Windsurf)
pip install markcrawl[mcp]
{
  "mcpServers": {
    "markcrawl": {
      "command": "python",
      "args": ["-m", "markcrawl.mcp_server"]
    }
  }
}

Tools: crawl_site, list_pages, read_page, search_pages, extract_data

LangChain Tool
pip install markcrawl[langchain]
from markcrawl.langchain import all_tools
from langchain_openai import ChatOpenAI
from langchain.agents import initialize_agent, AgentType

agent = initialize_agent(tools=all_tools, llm=ChatOpenAI(model="gpt-4o-mini"),
                         agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION)
agent.run("Crawl docs.example.com and summarize their auth guide")
OpenClaw Skill (WhatsApp, Telegram, Slack)
npx clawhub install markcrawl-skill

See AIMLPM/markcrawl-clawhub-skill.

LLM assistant prompt

Copy the system prompt from docs/LLM_PROMPT.md into any LLM to get an assistant that generates correct MarkCrawl commands.

When NOT to use MarkCrawl

  • Sites behind login/auth — no cookie or session support
  • Aggressive bot protection (Cloudflare, Akamai) — no anti-bot evasion
  • Millions of pages — designed for hundreds to low thousands; use Scrapy for scale
  • PDF content — HTML only (PDF support is on the roadmap)
  • JavaScript SPAs — add markcrawl[js] and use --render-js for React/Vue/Angular
  • Infinite-scroll pages--render-js renders the initial page load but does not scroll; you'll get the first screenful of content (e.g., ~28 of 82 YouTube videos). For complete listings, combine with the platform's API or RSS feed (e.g., YouTube's /feeds/videos.xml?channel_id=...)

Architecture

MarkCrawl is a web crawler. The optional layers (extraction, upload, agents) are separate add-ons that work with the crawler's output.

CORE (free, no API keys)              OPTIONAL ADD-ONS
┌──────────────────────────┐
│ 1. Discover URLs         │          markcrawl[extract]  — LLM field extraction
│    (sitemap or links)    │          markcrawl[upload]   — Supabase/pgvector RAG
│ 2. Fetch & clean HTML    │          markcrawl[js]       — Playwright JS rendering
│ 3. Write Markdown + JSONL│          markcrawl[mcp]      — MCP server for agents
│    + auto-citation       │          markcrawl[langchain] — LangChain tools
└──────────────────────────┘

For internals, see docs/ARCHITECTURE.md.

Extending MarkCrawl

from markcrawl import crawl

result = crawl("https://example.com", out_dir="./output")
print(f"Saved {result.pages_saved} pages")
# Process output in your own pipeline
import json
with open(result.index_file) as f:
    for line in f:
        page = json.loads(line)
        your_db.insert(page)  # Pinecone, Weaviate, Elasticsearch, etc.
# Use individual components
from markcrawl import chunk_text
from markcrawl.extract import LLMClient, extract_fields

See docs/ARCHITECTURE.md for the full module map and extensibility guide.

Cost

The core crawler is free. Two optional features have API costs:

Feature Cost When
Structured extraction ~$0.01-0.03 per page markcrawl-extract
Supabase upload ~$0.0001 per page markcrawl-upload

Setting up API keys

Only needed for extraction and upload. The core crawler requires no keys.

# .env — in your working directory
OPENAI_API_KEY="sk-..."           # extraction (--provider openai) + upload
ANTHROPIC_API_KEY="sk-ant-..."    # extraction (--provider anthropic)
GEMINI_API_KEY="AI..."            # extraction (--provider gemini)
XAI_API_KEY="xai-..."             # extraction (--provider grok)
SUPABASE_URL="https://..."        # upload
SUPABASE_KEY="eyJ..."             # upload (service-role key)
source .env
Project structure
.
├── README.md
├── LICENSE
├── PRIVACY.md
├── SECURITY.md
├── CONTRIBUTING.md
├── CODE_OF_CONDUCT.md
├── Dockerfile
├── Makefile
├── glama.json
├── pyproject.toml
├── requirements.txt
├── .github/
│   ├── pull_request_template.md
│   └── workflows/
│       ├── ci.yml
│       └── publish.yml
├── docs/
│   ├── ARCHITECTURE.md
│   ├── LLM_PROMPT.md
│   ├── MCP_SUBMISSION.md
│   ├── RAG_RETRIEVAL_RESEARCH.md
│   └── SUPABASE.md
├── tests/
│   ├── __init__.py
│   ├── test_chunker.py
│   ├── test_core.py
│   ├── test_extract.py
│   └── test_upload.py
└── markcrawl/
    ├── __init__.py
    ├── cli.py
    ├── core.py               # orchestrator
    ├── fetch.py              # HTTP/Playwright fetching
    ├── robots.py             # robots.txt parsing
    ├── throttle.py           # adaptive rate limiting
    ├── state.py              # crawl state & resume
    ├── urls.py               # URL normalization & filtering
    ├── extract_content.py    # HTML → Markdown conversion
    ├── dedup.py              # cross-crawl deduplication
    ├── link_scorer.py        # link prioritization
    ├── chunker.py
    ├── exceptions.py
    ├── utils.py
    ├── extract.py            # LLM field extraction
    ├── extract_cli.py
    ├── upload.py
    ├── upload_cli.py
    ├── langchain.py
    └── mcp_server.py

Roadmap

  • Canonical URL support
  • PDF support
  • Authenticated crawling
  • Multi-provider embeddings
Shipped features
  • pip install markcrawl on PyPI
  • 647 automated tests + GitHub Actions CI (Python 3.10-3.13) + ruff linting
  • Markdown and plain text output with auto-citation
  • Sitemap-first crawling with robots.txt compliance
  • Text chunking with configurable overlap + semantic chunking
  • Supabase/pgvector upload for RAG
  • JavaScript rendering via Playwright
  • Concurrent fetching and proxy support
  • Resume interrupted crawls + auto-resume
  • LLM extraction (OpenAI, Claude, Gemini, Grok) with auto-field discovery
  • MCP server, LangChain tools, OpenClaw skill
  • Image alt text preservation
  • Python API (result.pages)
  • Page-type extraction and content-region heuristics
  • Multiple extraction backends (default, trafilatura, ensemble, ReaderLM-v2)
  • Cross-crawl deduplication (--cross-dedup)
  • Link prioritization by predicted content yield (--prioritize-links)
  • Smart sampling of templated URL clusters (--smart-sample)
  • URL path filtering (--include-path, --exclude-path) and dry-run preview

Project info

  • Contributing — see CONTRIBUTING.md. If you used an LLM to generate code, include the prompt in your PR.
  • Security — see SECURITY.md for the disclosure policy.
  • Privacy — MarkCrawl runs locally. No telemetry, no analytics, no data sent anywhere. See PRIVACY.md.
  • License — MIT. See LICENSE.