webclaw

<p align="center"> <a href="https://webclaw.io"> <img src=".github/banner.png" alt="webclaw" width="700" /> </a> </p>

<h3 align="center"> The fastest web scraper for AI agents.<br/> _{67% fewer tokens. Sub-millisecond extraction. Zero browser overhead.} </h3>

<p align="center"> <a href="https://webclaw.io"><img src="https://img.shields.io/badge/website-webclaw.io-212529?style=flat-square" alt="Website" /></a> <a href="https://webclaw.io/docs"><img src="https://img.shields.io/badge/docs-webclaw.io%2Fdocs-212529?style=flat-square" alt="Docs" /></a> <a href="https://github.com/0xMassi/webclaw/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-MIT-212529?style=flat-square" alt="License" /></a> <a href="https://www.npmjs.com/package/create-webclaw"><img src="https://img.shields.io/npm/v/create-webclaw?style=flat-square&label=npx%20create-webclaw&color=212529" alt="npm" /></a> <a href="https://github.com/0xMassi/webclaw/stargazers"><img src="https://img.shields.io/github/stars/0xMassi/webclaw?style=flat-square&color=212529" alt="Stars" /></a> </p>

---

Your AI agent calls fetch() and gets a 403. Or 142KB of raw HTML that burns through your token budget. webclaw fixes both.

It extracts clean, structured content from any URL using Chrome-level TLS fingerprinting — no headless browser, no Selenium, no Puppeteer. Output is optimized for LLMs: 67% fewer tokens than raw HTML, with metadata, links, and images preserved.

                     Raw HTML                          webclaw
┌──────────────────────────────────┐    ┌──────────────────────────────────┐
│ <div class="ad-wrapper">         │    │ # Breaking: AI Breakthrough      │
│ <nav class="global-nav">         │    │                                  │
│ <script>window.__NEXT_DATA__     │    │ Researchers achieved 94%         │
│ ={...8KB of JSON...}</script>    │    │ accuracy on cross-domain         │
│ <div class="social-share">       │    │ reasoning benchmarks.            │
│ <button>Tweet</button>           │    │                                  │
│ <footer class="site-footer">     │    │ ## Key Findings                  │
│ <!-- 142,847 characters -->      │    │ - 3x faster inference            │
│                                  │    │ - Open-source weights            │
│         4,820 tokens             │    │         1,590 tokens             │
└──────────────────────────────────┘    └──────────────────────────────────┘

---

Get Started (30 seconds)

For AI agents (Claude, Cursor, Windsurf, VS Code)

npx create-webclaw

Auto-detects your AI tools, downloads the MCP server, and configures everything. One command.

CLI

# From source
git clone https://github.com/0xMassi/webclaw && cd webclaw
cargo build --release

# Or via Docker
docker run --rm ghcr.io/0xmassi/webclaw https://example.com

Docker Compose (with Ollama for LLM features)

cp env.example .env
docker compose up -d

---

Why webclaw?

	webclaw	Firecrawl	Trafilatura	Readability
Extraction accuracy	95.1%	—	80.6%	83.5%
Token efficiency	-67%	—	-55%	-51%
Speed (100KB page)	3.2ms	~500ms	18.4ms	8.7ms
TLS fingerprinting	Yes	No	No	No
Self-hosted	Yes	No	Yes	Yes
MCP (Claude/Cursor)	Yes	No	No	No
No browser required	Yes	No	Yes	Yes
Cost	Free	$$$$	Free	Free

Choose webclaw if you want fast local extraction, LLM-optimized output, and native AI agent integration.

---

What it looks like

$ webclaw https://stripe.com -f llm

> URL: https://stripe.com
> Title: Stripe | Financial Infrastructure for the Internet
> Language: en
> Word count: 847

# Stripe | Financial Infrastructure for the Internet

Stripe is a suite of APIs powering online payment processing
and commerce solutions for internet businesses of all sizes.

## Products
- Payments — Accept payments online and in person
- Billing — Manage subscriptions and invoicing
- Connect — Build a marketplace or platform
...

$ webclaw https://github.com --brand

{
  "name": "GitHub",
  "colors": [{"hex": "#59636E", "usage": "Primary"}, ...],
  "fonts": ["Mona Sans", "ui-monospace"],
  "logos": [{"url": "https://github.githubassets.com/...", "kind": "svg"}]
}

$ webclaw https://docs.rust-lang.org --crawl --depth 2 --max-pages 50

Crawling... 50/50 pages extracted
---
# Page 1: https://docs.rust-lang.org/
...
# Page 2: https://docs.rust-lang.org/book/
...

---

MCP Server — 10 tools for AI agents

webclaw ships as an MCP server that plugs into Claude Desktop, Claude Code, Cursor, Windsurf, OpenCode, Antigravity, Codex CLI, and any MCP-compatible client.

npx create-webclaw    # auto-detects and configures everything

Or manual setup — add to your Claude Desktop config:

{
  "mcpServers": {
    "webclaw": {
      "command": "~/.webclaw/webclaw-mcp"
    }
  }
}

Then in Claude: "Scrape the top 5 results for 'web scraping tools' and compare their pricing" — it just works.

Available tools

Tool	Description	Requires API key?
scrape	Extract content from any URL	No
crawl	Recursive site crawl	No
map	Discover URLs from sitemaps	No
batch	Parallel multi-URL extraction	No
extract	LLM-powered structured extraction	No (needs Ollama)
summarize	Page summarization	No (needs Ollama)
diff	Content change detection	No
brand	Brand identity extraction	No
search	Web search + scrape results	Yes
research	Deep multi-source research	Yes

8 of 10 tools work locally — no account, no API key, fully private.

---

Features

Extraction

• Readability scoring — multi-signal content detection (text density, semantic tags, link ratio)
• Noise filtering — strips nav, footer, ads, modals, cookie banners (Tailwind-safe)
• Data island extraction — catches React/Next.js JSON payloads, JSON-LD, hydration data
• YouTube metadata — structured data from any YouTube video
• PDF extraction — auto-detected via Content-Type
• 5 output formats — markdown, text, JSON, LLM-optimized, HTML

Content control

webclaw URL --include "article, .content"       # CSS selector include
webclaw URL --exclude "nav, footer, .sidebar"    # CSS selector exclude
webclaw URL --only-main-content                  # Auto-detect main content

Crawling

webclaw URL --crawl --depth 3 --max-pages 100   # BFS same-origin crawl
webclaw URL --crawl --sitemap                    # Seed from sitemap
webclaw URL --map                                # Discover URLs only

LLM features (Ollama / OpenAI / Anthropic)

webclaw URL --summarize                          # Page summary
webclaw URL --extract-prompt "Get all prices"    # Natural language extraction
webclaw URL --extract-json '{"type":"object"}'   # Schema-enforced extraction

Change tracking

webclaw URL -f json > snap.json                  # Take snapshot
webclaw URL --diff-with snap.json                # Compare later

Brand extraction

webclaw URL --brand                              # Colors, fonts, logos, OG image

Proxy rotation

webclaw URL --proxy http://user:pass@host:port   # Single proxy
webclaw URLs --proxy-file proxies.txt            # Pool rotation

---

Benchmarks

All numbers from real tests on 50 diverse pages. See benchmarks/ for methodology and reproduction instructions.

Extraction quality

Accuracy      webclaw     ███████████████████ 95.1%
              readability ████████████████▋   83.5%
              trafilatura ████████████████    80.6%
              newspaper3k █████████████▎      66.4%

Noise removal webclaw     ███████████████████ 96.1%
              readability █████████████████▊  89.4%
              trafilatura ██████████████████▏ 91.2%
              newspaper3k ███████████████▎    76.8%

Speed (pure extraction, no network)

10KB page     webclaw     ██                   0.8ms
              readability █████                2.1ms
              trafilatura ██████████           4.3ms

100KB page    webclaw     ██                   3.2ms
              readability █████                8.7ms
              trafilatura ██████████           18.4ms

Token efficiency (feeding to Claude/GPT)

Format	Tokens	vs Raw HTML
Raw HTML	4,820	baseline
readability	2,340	-51%
trafilatura	2,180	-55%
webclaw llm	1,590	-67%

Crawl speed

Concurrency	webclaw	Crawl4AI	Scrapy
5	9.8 pg/s	5.2 pg/s	7.1 pg/s
10	18.4 pg/s	8.7 pg/s	12.3 pg/s
20	32.1 pg/s	14.2 pg/s	21.8 pg/s

---

Architecture

webclaw/
  crates/
    webclaw-core     Pure extraction engine. Zero network deps. WASM-safe.
    webclaw-fetch    HTTP client + TLS fingerprinting. Crawler. Batch ops.
    webclaw-llm      LLM provider chain (Ollama -> OpenAI -> Anthropic)
    webclaw-pdf      PDF text extraction
    webclaw-mcp      MCP server (10 tools for AI agents)
    webclaw-cli      CLI binary

webclaw-core takes raw HTML as a &str and returns structured output. No I/O, no network, no allocator tricks. Can compile to WASM.

---

Configuration

Variable	Description
WEBCLAW_API_KEY	Cloud API key (enables bot bypass, JS rendering, search, research)
OLLAMA_HOST	Ollama URL for local LLM features (default: http://localhost:11434)
OPENAI_API_KEY	OpenAI API key for LLM features
ANTHROPIC_API_KEY	Anthropic API key for LLM features
WEBCLAW_PROXY	Single proxy URL
WEBCLAW_PROXY_FILE	Path to proxy pool file

---

Cloud API (optional)

For bot-protected sites, JS rendering, and advanced features, webclaw offers a hosted API at webclaw.io.

The CLI and MCP server work locally first. Cloud is used as a fallback when:

• A site has bot protection (Cloudflare, DataDome, WAF)
• A page requires JavaScript rendering
• You use search or research tools

export WEBCLAW_API_KEY=wc_your_key

# Automatic: tries local first, cloud on bot detection
webclaw https://protected-site.com

# Force cloud
webclaw --cloud https://spa-site.com

SDKs

npm install @webclaw/sdk                  # TypeScript/JavaScript
pip install webclaw                        # Python
go get github.com/0xMassi/webclaw-go      # Go

---

Use cases

• AI agents — Give Claude/Cursor/GPT real-time web access via MCP
• Research — Crawl documentation, competitor sites, news archives
• Price monitoring — Track changes with --diff-with snapshots
• Training data — Prepare web content for fine-tuning with token-optimized output
• Content pipelines — Batch extract + summarize in CI/CD
• Brand intelligence — Extract visual identity from any website

---

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

• Good first issues
• Architecture docs

License

MIT — use it however you want.