crw

The web scraper built for AI agents. Single binary. Zero config.

Works with: Claude Code · Cursor · Windsurf · Cline · Copilot · Continue.dev · Codex

MCP Integration • Installation • API Reference • Cloud • JS Rendering • Configuration • Discord

English | 中文

Don't want to self-host? fastcrw.com is the managed cloud — global proxy network, auto-scaling, dashboard, and API keys. Same Firecrawl-compatible API. Get 500 free credits →

CRW is the open-source web scraper built for AI agents. Built-in MCP server (stdio + HTTP), single binary, ~6 MB idle RAM. Give Claude Code, Cursor, or any MCP client web scraping superpowers in 30 seconds. Firecrawl-compatible API — 5.5x faster, 75x less memory, 92% coverage on 1K real-world URLs.

Built-in MCP server. Single binary. No Redis. No Node.js.

# npm (zero install):
npx crw-mcp

# Python:
pip install crw

# Direct binary (no package manager):
curl -fsSL https://github.com/us/crw/releases/latest/download/crw-mcp-darwin-arm64.tar.gz | tar xz
# Replace darwin-arm64 with your platform: darwin-x64, linux-x64, linux-arm64, win32-x64, win32-arm64

# Cargo:
cargo install crw-mcp

# Docker:
docker run -i ghcr.io/us/crw crw-mcp

Listed on the MCP Registry

What's New

0.3.0 (2026-04-02)

Features

• renderer: escalate to JS renderer on HTTP 401/403 responses (88c8aa5)

Bug Fixes

• use GitHub latest release instead of pinned version for binary download (4afcb1a)

0.2.1 (2026-03-28)

Bug Fixes

• make crw-mcp npm wrapper executable (576a9eb)
• use latest tag in server.json OCI identifier (7ec3b82)

0.2.0 (2026-03-28)

Features

• add MCP Registry support for official server discovery (154b9f5)

Full changelog →

Why CRW?

CRW gives you Firecrawl's API with a fraction of the resource usage. No runtime dependencies, no Redis, no Node.js — just a single binary you can deploy anywhere.

CRW vs Firecrawl — Tested on Firecrawl scrape-content-dataset-v1 (1,000 real-world URLs, JS rendering enabled):

• CRW covers 92% of URLs vs Firecrawl's 77.2% — 15 percentage points higher
• CRW is 5.5x faster on average (833ms vs 4,600ms)
• CRW uses ~75x less RAM at idle (6.6 MB vs ~500 MB+)
• Firecrawl requires 5 containers (Node.js, Redis, PostgreSQL, RabbitMQ, Playwright) — CRW is a single binary

Crawl4AI vs Firecrawl vs Spider — Independent benchmark by Spider.cloud:

Firecrawl independent review — Scrapeway benchmark: 64.3% success rate, $5.11/1K scrapes, 0% on LinkedIn/Twitter.

Resource comparison:

Features

• 🔧 MCP server — built-in stdio + HTTP transport for Claude Code, Cursor, Windsurf, and any MCP client
• 🔌 Firecrawl-compatible API — same endpoint family and familiar request/response ergonomics
• 📄 6 output formats — markdown, HTML, cleaned HTML, raw HTML, plain text, links, structured JSON
• 🤖 LLM structured extraction — send a JSON schema, get validated structured data back (Anthropic tool_use + OpenAI function calling)
• 🌐 JS rendering — auto-detect SPAs with shell heuristics, render via LightPanda, Playwright, or Chrome (CDP)
• 🕷️ BFS crawler — async crawl with rate limiting, robots.txt, sitemap support, concurrent jobs
• 🔒 Security — SSRF protection (private IPs, cloud metadata, IPv6), constant-time auth, dangerous URI filtering
• 🐳 Docker ready — multi-stage build with LightPanda sidecar
• 🎯 CSS selector & XPath — extract specific DOM elements before Markdown conversion
• ✂️ Chunking & filtering — split content into topic/sentence/regex chunks; rank by BM25 or cosine similarity
• 🕵️ Stealth mode — browser-like UA rotation and header injection to reduce bot detection
• 🌐 Per-request proxy — override the global proxy per scrape request

Cloud vs Self-Hosted

Both use the same Firecrawl-compatible API — your code works with either. Switch between self-hosted and cloud by changing the base URL.

Quick Start

MCP (AI agents — recommended):

claude mcp add crw -- npx crw-mcp

That's it. Claude Code now has crw_scrape, crw_crawl, crw_map tools. For Cursor, Windsurf, Cline, and other MCP clients, see MCP Server.

CLI (no server needed):

cargo install crw-cli
crw https://example.com

Self-hosted server:

cargo install crw-server
crw-server

Enable JS rendering (optional):

crw-server setup

This downloads LightPanda and creates a config.local.toml for JS rendering. See JS Rendering for details.

Cloud (no setup):

curl -X POST https://fastcrw.com/api/v1/scrape \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com"}'

Get your API key at fastcrw.com — 500 free credits included.

Docker:

docker run -p 3000:3000 ghcr.io/us/crw:latest

Docker Compose (with JS rendering):

docker compose up

Scrape a page:

curl -X POST http://localhost:3000/v1/scrape \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com"}'

{
  "success": true,
  "data": {
    "markdown": "# Example Domain\nThis domain is for use in ...",
    "metadata": {
      "title": "Example Domain",
      "sourceURL": "https://example.com",
      "statusCode": 200,
      "elapsedMs": 32
    }
  }
}

Use Cases

• RAG pipelines — crawl websites and extract structured data for vector databases
• AI agents — give Claude Code or Claude Desktop web scraping tools via MCP
• Content monitoring — periodic crawl with LLM extraction to track changes
• Data extraction — combine CSS selectors + LLM to extract any schema from any page
• Web archiving — full-site BFS crawl to markdown

API Endpoints

LLM Structured Extraction

Send a JSON schema with your scrape request and CRW returns validated structured data using LLM function calling.

curl -X POST http://localhost:3000/v1/scrape \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://example.com/product",
    "formats": ["json"],
    "jsonSchema": {
      "type": "object",
      "properties": {
        "name": { "type": "string" },
        "price": { "type": "number" }
      },
      "required": ["name", "price"]
    }
  }'

• Anthropic — uses tool_use with input_schema for extraction
• OpenAI — uses function calling with parameters schema
• Validation — LLM output is validated against your JSON schema before returning

Configure the LLM provider in your config:

[extraction.llm]
provider = "anthropic"        # "anthropic" or "openai"
api_key = "sk-..."            # or CRW_EXTRACTION__LLM__API_KEY env var
model = "claude-sonnet-4-20250514"

MCP Server

CRW works as an MCP tool server for any AI assistant that supports MCP. It provides 4 tools: crw_scrape, crw_crawl, crw_check_crawl_status, crw_map.

Also available on the MCP Registry

Install:

# npm (zero install):
npx crw-mcp

# Python:
pip install crw

# Direct binary (no package manager):
curl -fsSL https://github.com/us/crw/releases/latest/download/crw-mcp-darwin-arm64.tar.gz | tar xz
# Replace darwin-arm64 with your platform: darwin-x64, linux-x64, linux-arm64, win32-x64, win32-arm64

# Cargo:
cargo install crw-mcp

# Docker:
docker run -i ghcr.io/us/crw crw-mcp

Claude Code

claude mcp add crw -- npx crw-mcp

Claude Desktop

Edit your config file:

{
  "mcpServers": {
    "crw": {
      "command": "npx",
      "args": ["crw-mcp"]
    }
  }
}

Cursor

Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "crw": {
      "command": "npx",
      "args": ["crw-mcp"]
    }
  }
}

Windsurf

Edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "crw": {
      "command": "npx",
      "args": ["crw-mcp"]
    }
  }
}

Cline (VS Code)

{
  "mcpServers": {
    "crw": {
      "command": "npx",
      "args": ["crw-mcp"],
      "alwaysAllow": ["crw_scrape", "crw_map"],
      "disabled": false
    }
  }
}

Continue.dev (VS Code / JetBrains)

Edit ~/.continue/config.yaml:

mcpServers:
  - name: crw
    command: npx
    args:
      - crw-mcp

OpenAI Codex CLI

Edit ~/.codex/config.toml:

[mcp_servers.crw]
command = "npx"
args = ["crw-mcp"]

Other MCP Clients

Any MCP-compatible client can connect to CRW using the standard JSON format:

{
  "mcpServers": {
    "crw": {
      "command": "npx",
      "args": ["crw-mcp"]
    }
  }
}

Tip: The stdio binary (crw-mcp) works with any client. For clients that support HTTP transport, use http://localhost:3000/mcp directly — no binary needed.

See the full MCP setup guide for detailed instructions, auth configuration, and platform comparison.

JS Rendering

CRW auto-detects SPAs by analyzing the initial HTML response for shell heuristics (empty body, framework markers). When a SPA is detected, it renders the page via a headless browser.

Quick setup (recommended):

crw-server setup

This automatically downloads the LightPanda binary to ~/.local/bin/ and creates a config.local.toml with the correct renderer settings. Then start LightPanda and CRW:

lightpanda serve --host 127.0.0.1 --port 9222 &
crw-server

Supported renderers:

Renderer mode is configured via renderer.mode: auto (default), lightpanda, playwright, chrome, or none.

With Docker Compose, LightPanda runs as a sidecar — no extra setup needed:

docker compose up

Architecture

┌─────────────────────────────────────────────┐
│                 crw-server                  │
│         Axum HTTP API + Auth + MCP          │
├──────────┬──────────┬───────────────────────┤
│ crw-crawl│crw-extract│    crw-renderer      │
│ BFS crawl│ HTML→MD   │  HTTP + CDP(WS)      │
│ robots   │ LLM/JSON  │  LightPanda/Chrome   │
│ sitemap  │ clean/read│  auto-detect SPA     │
├──────────┴──────────┴───────────────────────┤
│                 crw-core                    │
│        Types, Config, Errors                │
└─────────────────────────────────────────────┘

Configuration

CRW uses layered TOML configuration with environment variable overrides:

1. config.default.toml — built-in defaults
2. config.local.toml — local overrides (or set CRW_CONFIG=myconfig)
3. Environment variables — CRW_ prefix with __ separator (e.g. CRW_SERVER__PORT=8080)

[server]
host = "0.0.0.0"
port = 3000
rate_limit_rps = 10        # Max requests/second (global). 0 = unlimited.

[renderer]
mode = "auto"  # auto | lightpanda | playwright | chrome | none

[crawler]
max_concurrency = 10
requests_per_second = 10.0
respect_robots_txt = true

[auth]
# api_keys = ["fc-key-1234"]

See full configuration reference for all options.

Integration Examples

Python:

import requests

response = requests.post("http://localhost:3000/v1/scrape", json={
    "url": "https://example.com",
    "formats": ["markdown", "links"]
})
data = response.json()["data"]
print(data["markdown"])

Node.js:

const response = await fetch("http://localhost:3000/v1/scrape", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
    url: "https://example.com",
    formats: ["markdown", "links"]
  })
});
const { data } = await response.json();
console.log(data.markdown);

CrewAI (crewai-crw — CRW tools for CrewAI):

pip install crewai-crw

from crewai import Agent, Task, Crew
from crewai_crw import CrwScrapeWebsiteTool, CrwCrawlWebsiteTool, CrwMapWebsiteTool

scrape_tool = CrwScrapeWebsiteTool()  # uses localhost:3000 by default

researcher = Agent(
    role="Web Researcher",
    goal="Research and summarize information from websites",
    backstory="Expert at extracting key information from web pages",
    tools=[scrape_tool],
)

task = Task(
    description="Scrape https://example.com and summarize the content",
    expected_output="A summary of the page content",
    agent=researcher,
)

crew = Crew(agents=[researcher], tasks=[task])
result = crew.kickoff()

LangChain (langchain-crw — CRW document loader):

pip install langchain-crw

from langchain_crw import CrwLoader

# Self-hosted (default: localhost:3000)
loader = CrwLoader(url="https://example.com", mode="scrape")
docs = loader.load()
print(docs[0].page_content)  # clean markdown

# Cloud (fastcrw.com)
loader = CrwLoader(
    url="https://example.com",
    api_url="https://fastcrw.com/api",
    api_key="crw_live_...",
    mode="crawl",
    params={"max_depth": 3, "max_pages": 50},
)
docs = loader.load()

More integrations (PRs pending): Flowise · Agno · n8n · All integrations

Docker

Pre-built image from GHCR:

docker pull ghcr.io/us/crw:latest
docker run -p 3000:3000 ghcr.io/us/crw:latest

Docker Compose (with JS rendering sidecar):

docker compose up

This starts CRW on port 3000 with LightPanda as a JS rendering sidecar on port 9222. CRW auto-connects to LightPanda for SPA rendering.

Benchmark

Tested on Firecrawl's scrape-content-dataset-v1 (1,000 real-world URLs, JS rendering enabled):

CRW hits a sweet spot: near-Spider speed with Firecrawl API compatibility and ~75x less RAM. Unlike Crawl4AI (Python + Playwright), CRW ships as a single Rust binary with no runtime dependencies.

Run the benchmark yourself:

uv pip install datasets aiohttp
uv run python bench/run_bench.py

Crates

Crate	Description
crw-core	Core types, config, and error handling
crw-renderer	HTTP + CDP browser rendering engine
crw-extract	HTML → markdown/plaintext extraction
crw-crawl	Async BFS crawler with robots.txt & sitemap
crw-server	Axum API server (Firecrawl-compatible)
crw-cli	Standalone CLI (crw binary, no server needed)
crw-mcp	MCP stdio proxy binary

See docs/crates.md for usage examples and cargo add instructions.

Documentation

Full documentation: docs/index.md

• Getting Started
• Configuration
• API Reference
• MCP Server
• JS Rendering
• Architecture

Security

CRW includes built-in protections against common web scraping attack vectors:

• SSRF protection — all URL inputs (REST API + MCP) are validated against private/internal networks:
  • Loopback (127.0.0.0/8, ::1, localhost)
  • Private IPs (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16)
  • Link-local / cloud metadata (169.254.0.0/16 — blocks AWS/GCP metadata endpoints)
  • IPv6 mapped addresses (::ffff:127.0.0.1), link-local (fe80::), ULA (fc00::/7)
  • Non-HTTP schemes (file://, ftp://, gopher://, data:)
• Auth — optional Bearer token with constant-time comparison (no length or key-index leakage)
• robots.txt — respects Allow/Disallow with wildcard patterns (*, $) and RFC 9309 specificity
• Rate limiting — configurable per-second request cap with token-bucket algorithm (returns 429 with error_code)
• Resource limits — max body size (1 MB), max crawl depth (10), max pages (1000), max discovered URLs (5000)

Community

Don't be shy — join us on Discord! Ask questions, share what you're building, report bugs, or just hang out.

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

1. Fork the repository
2. Install pre-commit hooks: make hooks
3. Create your feature branch (git checkout -b feat/my-feature)
4. Commit your changes (git commit -m 'feat: add my feature')
5. Push to the branch (git push origin feat/my-feature)
6. Open a Pull Request

The pre-commit hook runs the same checks as CI (cargo fmt, cargo clippy, cargo test). You can also run them manually with make check.

License

CRW is open-source under AGPL-3.0. For a managed version without AGPL obligations, see fastcrw.com.