MCP Server for Crawl4AI

Note: Tested with Crawl4AI version 0.7.4

TypeScript implementation of an MCP server for Crawl4AI. Provides tools for web crawling, content extraction, and browser automation.

Table of Contents

• Prerequisites
• Quick Start
• Configuration
• Client-Specific Instructions
• Available Tools
  • 1. get_markdown
  • 2. capture_screenshot
  • 3. generate_pdf
  • 4. execute_js
  • 5. batch_crawl
  • 6. smart_crawl
  • 7. get_html
  • 8. extract_links
  • 9. crawl_recursive
  • 10. parse_sitemap
  • 11. crawl
  • 12. manage_session
  • 13. extract_with_llm
• Advanced Configuration
• Development
• License

Prerequisites

• Node.js 18+ and npm
• A running Crawl4AI server

Quick Start

1. Start the Crawl4AI server (for example, local docker)

docker run -d -p 11235:11235 --name crawl4ai --shm-size=1g unclecode/crawl4ai:0.7.4

2. Add to your MCP client

This MCP server works with any MCP-compatible client (Claude Desktop, Claude Code, Cursor, LMStudio, etc.).

Using npx (Recommended)

{
  "mcpServers": {
    "crawl4ai": {
      "command": "npx",
      "args": ["mcp-crawl4ai-ts"],
      "env": {
        "CRAWL4AI_BASE_URL": "http://localhost:11235"
      }
    }
  }
}

Using local installation

{
  "mcpServers": {
    "crawl4ai": {
      "command": "node",
      "args": ["/path/to/mcp-crawl4ai-ts/dist/index.js"],
      "env": {
        "CRAWL4AI_BASE_URL": "http://localhost:11235"
      }
    }
  }
}

With all optional variables

{
  "mcpServers": {
    "crawl4ai": {
      "command": "npx",
      "args": ["mcp-crawl4ai-ts"],
      "env": {
        "CRAWL4AI_BASE_URL": "http://localhost:11235",
        "CRAWL4AI_API_KEY": "your-api-key",
        "SERVER_NAME": "custom-name",
        "SERVER_VERSION": "1.0.0"
      }
    }
  }
}

Configuration

Environment Variables

# Required
CRAWL4AI_BASE_URL=http://localhost:11235

# Optional - Server Configuration
CRAWL4AI_API_KEY=          # If your server requires auth
SERVER_NAME=crawl4ai-mcp   # Custom name for the MCP server
SERVER_VERSION=1.0.0       # Custom version

Client-Specific Instructions

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json

Claude Code

claude mcp add crawl4ai -e CRAWL4AI_BASE_URL=http://localhost:11235 -- npx mcp-crawl4ai-ts

Other MCP Clients

Consult your client's documentation for MCP server configuration. The key details:

• Command: npx mcp-crawl4ai-ts or node /path/to/dist/index.js
• Required env: CRAWL4AI_BASE_URL
• Optional env: CRAWL4AI_API_KEY, SERVER_NAME, SERVER_VERSION

Available Tools

1. get_markdown - Extract content as markdown with filtering

{ 
  url: string,                              // Required: URL to extract markdown from
  filter?: 'raw'|'fit'|'bm25'|'llm',       // Filter type (default: 'fit')
  query?: string,                           // Query for bm25/llm filters
  cache?: string                            // Cache-bust parameter (default: '0')
}

Extracts content as markdown with various filtering options. Use 'bm25' or 'llm' filters with a query for specific content extraction.

2. capture_screenshot - Capture webpage screenshot

{ 
  url: string,                   // Required: URL to capture
  screenshot_wait_for?: number   // Seconds to wait before screenshot (default: 2)
}

Returns base64-encoded PNG. Note: This is stateless - for screenshots after JS execution, use crawl with screenshot: true.

3. generate_pdf - Convert webpage to PDF

{ 
  url: string  // Required: URL to convert to PDF
}

Returns base64-encoded PDF. Stateless tool - for PDFs after JS execution, use crawl with pdf: true.

4. execute_js - Execute JavaScript and get return values

{ 
  url: string,                    // Required: URL to load
  scripts: string | string[]      // Required: JavaScript to execute
}

Executes JavaScript and returns results. Each script can use 'return' to get values back. Stateless - for persistent JS execution use crawl with js_code.

5. batch_crawl - Crawl multiple URLs concurrently

{ 
  urls: string[],           // Required: List of URLs to crawl
  max_concurrent?: number,  // Parallel request limit (default: 5)
  remove_images?: boolean,  // Remove images from output (default: false)
  bypass_cache?: boolean,   // Bypass cache for all URLs (default: false)
  configs?: Array<{         // Optional: Per-URL configurations (v3.0.0+)
    url: string,
    [key: string]: any      // Any crawl parameters for this specific URL
  }>
}

Efficiently crawls multiple URLs in parallel. Each URL gets a fresh browser instance. With configs array, you can specify different parameters for each URL.

6. smart_crawl - Auto-detect and handle different content types

{ 
  url: string,            // Required: URL to crawl
  max_depth?: number,     // Maximum depth for recursive crawling (default: 2)
  follow_links?: boolean, // Follow links in content (default: true)
  bypass_cache?: boolean  // Bypass cache (default: false)
}

Intelligently detects content type (HTML/sitemap/RSS) and processes accordingly.

7. get_html - Get sanitized HTML for analysis

{ 
  url: string  // Required: URL to extract HTML from
}

Returns preprocessed HTML optimized for structure analysis. Use for building schemas or analyzing patterns.

8. extract_links - Extract and categorize page links

{ 
  url: string,          // Required: URL to extract links from
  categorize?: boolean  // Group by type (default: true)
}

Extracts all links and groups them by type: internal, external, social media, documents, images.

9. crawl_recursive - Deep crawl website following links

{ 
  url: string,              // Required: Starting URL
  max_depth?: number,       // Maximum depth to crawl (default: 3)
  max_pages?: number,       // Maximum pages to crawl (default: 50)
  include_pattern?: string, // Regex pattern for URLs to include
  exclude_pattern?: string  // Regex pattern for URLs to exclude
}

Crawls a website following internal links up to specified depth. Returns content from all discovered pages.

10. parse_sitemap - Extract URLs from XML sitemaps

{ 
  url: string,              // Required: Sitemap URL (e.g., /sitemap.xml)
  filter_pattern?: string   // Optional: Regex pattern to filter URLs
}

Extracts all URLs from XML sitemaps. Supports regex filtering for specific URL patterns.

11. crawl - Advanced web crawling with full configuration

{
  url: string,                              // URL to crawl
  // Browser Configuration
  browser_type?: 'chromium'|'firefox'|'webkit'|'undetected',  // Browser engine (undetected = stealth mode)
  viewport_width?: number,                  // Browser width (default: 1080)
  viewport_height?: number,                 // Browser height (default: 600)
  user_agent?: string,                      // Custom user agent
  proxy_server?: string | {                 // Proxy URL (string or object format)
    server: string,
    username?: string,
    password?: string
  },
  proxy_username?: string,                  // Proxy auth (if using string format)
  proxy_password?: string,                  // Proxy password (if using string format)
  cookies?: Array<{name, value, domain}>,   // Pre-set cookies
  headers?: Record<string,string>,          // Custom headers
  
  // Crawler Configuration
  word_count_threshold?: number,            // Min words per block (default: 200)
  excluded_tags?: string[],                 // HTML tags to exclude
  remove_overlay_elements?: boolean,        // Remove popups/modals
  js_code?: string | string[],              // JavaScript to execute
  wait_for?: string,                        // Wait condition (selector or JS)
  wait_for_timeout?: number,                // Wait timeout (default: 30000)
  delay_before_scroll?: number,             // Pre-scroll delay
  scroll_delay?: number,                    // Between-scroll delay
  process_iframes?: boolean,                // Include iframe content
  exclude_external_links?: boolean,         // Remove external links
  screenshot?: boolean,                     // Capture screenshot
  pdf?: boolean,                           // Generate PDF
  session_id?: string,                      // Reuse browser session (only works with crawl tool)
  cache_mode?: 'ENABLED'|'BYPASS'|'DISABLED',  // Cache control
  
  // New in v3.0.0 (Crawl4AI 0.7.3/0.7.4)
  css_selector?: string,                    // CSS selector to filter content
  delay_before_return_html?: number,        // Delay in seconds before returning HTML
  include_links?: boolean,                  // Include extracted links in response
  resolve_absolute_urls?: boolean,          // Convert relative URLs to absolute
  
  // LLM Extraction (REST API only supports 'llm' type)
  extraction_type?: 'llm',                  // Only 'llm' extraction is supported via REST API
  extraction_schema?: object,               // Schema for structured extraction
  extraction_instruction?: string,          // Natural language extraction prompt
  extraction_strategy?: {                   // Advanced extraction configuration
    provider?: string,
    api_key?: string,
    model?: string,
    [key: string]: any
  },
  table_extraction_strategy?: {             // Table extraction configuration
    enable_chunking?: boolean,
    thresholds?: object,
    [key: string]: any
  },
  markdown_generator_options?: {            // Markdown generation options
    include_links?: boolean,
    preserve_formatting?: boolean,
    [key: string]: any
  },
  
  timeout?: number,                         // Overall timeout (default: 60000)
  verbose?: boolean                         // Detailed logging
}

12. manage_session - Unified session management

{ 
  action: 'create' | 'clear' | 'list',    // Required: Action to perform
  session_id?: string,                    // For 'create' and 'clear' actions
  initial_url?: string,                   // For 'create' action: URL to load
  browser_type?: 'chromium' | 'firefox' | 'webkit' | 'undetected'  // For 'create' action
}

Unified tool for managing browser sessions. Supports three actions:

• create: Start a persistent browser session
• clear: Remove a session from local tracking
• list: Show all active sessions

Examples:

// Create a new session
{ action: 'create', session_id: 'my-session', initial_url: 'https://example.com' }

// Clear a session
{ action: 'clear', session_id: 'my-session' }

// List all sessions
{ action: 'list' }

13. extract_with_llm - Extract structured data using AI

{ 
  url: string,          // URL to extract data from
  query: string         // Natural language extraction instructions
}

Uses AI to extract structured data from webpages. Returns results immediately without any polling or job management. This is the recommended way to extract specific information since CSS/XPath extraction is not supported via the REST API.

Advanced Configuration

For detailed information about all available configuration options, extraction strategies, and advanced features, please refer to the official Crawl4AI documentation:

• Crawl4AI Documentation
• Crawl4AI GitHub Repository

Changelog

See CHANGELOG.md for detailed version history and recent updates.

Development

Setup

# 1. Start the Crawl4AI server
docker run -d -p 11235:11235 --name crawl4ai --shm-size=1g unclecode/crawl4ai:latest

# 2. Install MCP server
git clone https://github.com/omgwtfwow/mcp-crawl4ai-ts.git
cd mcp-crawl4ai-ts
npm install
cp .env.example .env

# 3. Development commands
npm run dev    # Development mode
npm test       # Run tests
npm run lint   # Check code quality
npm run build  # Production build

# 4. Add to your MCP client (See "Using local installation")

Running Integration Tests

Integration tests require a running Crawl4AI server. Configure your environment:

# Required for integration tests
export CRAWL4AI_BASE_URL=http://localhost:11235
export CRAWL4AI_API_KEY=your-api-key  # If authentication is required

# Optional: For LLM extraction tests
export LLM_PROVIDER=openai/gpt-4o-mini
export LLM_API_TOKEN=your-llm-api-key
export LLM_BASE_URL=https://api.openai.com/v1  # If using custom endpoint

# Run integration tests (ALWAYS use the npm script; don't call `jest` directly)
npm run test:integration

# Run a single integration test file
npm run test:integration -- src/__tests__/integration/extract-links.integration.test.ts

> IMPORTANT: Do NOT run `npx jest` directly for integration tests. The npm script injects `NODE_OPTIONS=--experimental-vm-modules` which is required for ESM + ts-jest. Running Jest directly will produce `SyntaxError: Cannot use import statement outside a module` and hang.

Integration tests cover:

• Dynamic content and JavaScript execution
• Session management and cookies
• Content extraction (LLM-based only)
• Media handling (screenshots, PDFs)
• Performance and caching
• Content filtering
• Bot detection avoidance
• Error handling

Integration Test Checklist

1. Docker container healthy:

docker ps --filter name=crawl4ai --format '{{.Names}} {{.Status}}'
curl -sf http://localhost:11235/health || echo "Health check failed"

2. Env vars loaded (either exported or in .env): CRAWL4AI_BASE_URL (required), optional: CRAWL4AI_API_KEY, LLM_PROVIDER, LLM_API_TOKEN, LLM_BASE_URL.
3. Use npm run test:integration (never raw jest).
4. To target one file add it after -- (see example above).
5. Expect total runtime ~2–3 minutes; longer or immediate hang usually means missing NODE_OPTIONS or wrong Jest version.

Troubleshooting

| Symptom | Likely Cause | Fix | |---------|--------------|-----| | SyntaxError: Cannot use import statement outside a module | Ran jest directly without script flags | Re-run with npm run test:integration | | Hangs on first test (RUNS ...) | Missing experimental VM modules flag | Use npm script / ensure NODE_OPTIONS=--experimental-vm-modules | | Network timeouts | Crawl4AI container not healthy / DNS blocked | Restart container: docker restart <name> | | LLM tests skipped | Missing LLM_PROVIDER or LLM_API_TOKEN | Export required LLM vars | | New Jest major upgrade breaks tests | Version mismatch with ts-jest | Keep Jest 29.x unless ts-jest upgraded accordingly |

Version Compatibility Note

Current stack: [email protected] + [email protected] + ESM ("type": "module"). Updating Jest to 30+ requires upgrading ts-jest and revisiting jest.config.cjs. Keep versions aligned to avoid parse errors.

License

MIT