Crawl4AI-MCP

A Model Context Protocol server for web crawling powered by Crawl4AI

Overview

Crawl4AI-MCP is a Model Context Protocol server that gives AI systems access to the live web. Built on FastMCP v3 and Crawl4AI, it exposes 4 tools, 2 resources, and 3 prompts through the standardized MCP interface, backed by a lifespan-managed headless Chromium browser.

Only 2 runtime dependencies — fastmcp and crawl4ai.

[!TIP] Full documentation site →

Key Features

• Full MCP compliance via FastMCP v3 with tool annotations (readOnlyHint, destructiveHint, etc.)
• 4 focused tools centered on canonical scrape/crawl plus session lifecycle/artifacts
• 3 prompts for common LLM workflows (summarize, extract schema, compare pages)
• 2 resources exposing server configuration and version info
• Headless Chromium managed as a lifespan singleton (start once, reuse everywhere)
• Multiple transports — stdio (default) and Streamable HTTP
• LLM-optimized output — markdown, cleaned HTML, raw HTML, or plain text
• Canonical option groups for extraction, runtime, diagnostics, sessions, rendering, and traversal
• List and deep traversal in one crawl contract
• Session-aware workflows with explicit session close and artifact retrieval tools
• Auto browser setup — detects missing Playwright browsers and installs automatically

Installation

pip install crawl4ai-mcp
crawl4ai-mcp --setup       # one-time: installs Playwright browsers

uv add crawl4ai-mcp
crawl4ai-mcp --setup       # one-time: installs Playwright browsers

docker build -t crawl4ai-mcp .
docker run -p 8000:8000 crawl4ai-mcp

git clone https://github.com/wyattowalsh/crawl4ai-mcp.git
cd crawl4ai-mcp
uv sync --group dev
crawl4ai-mcp --setup

[!NOTE] The server auto-detects missing Playwright browsers on first startup and attempts to install them automatically. You can also run crawl4ai-mcp --setup or crawl4ai-setup manually at any time.

Quick Start

crawl4ai-mcp

crawl4ai-mcp --transport http --port 8000

{
  "mcpServers": {
    "crawl4ai": {
      "command": "crawl4ai-mcp",
      "args": ["--transport", "stdio"]
    }
  }
}

claude mcp add crawl4ai -- crawl4ai-mcp --transport stdio

npx @modelcontextprotocol/inspector crawl4ai-mcp

Tools

The canonical surface now exposes 4 tools:

scrape

Scrape one URL or a bounded list of URLs (up to 20) with a single canonical envelope response.

• Input: targets (str | list[str]) and optional grouped options
• Supports extraction (schema, extraction_mode), runtime controls, diagnostics, session settings, render settings, and artifact capture
• Returns canonical JSON envelope with schema_version, tool, ok, data/items, meta, warnings, error

crawl

Crawl with canonical traversal controls.

• options.traversal.mode="list" for bounded list traversal
• options.traversal.mode="deep" for recursive BFS/DFS traversal from a single seed
• Shares scrape option groups plus traversal options in the same envelope shape

close_session

Close a stateful session created via options.session.session_id.

get_artifact

Retrieve artifact metadata/content captured during scrape or crawl when options.conversion.capture_artifacts is enabled.

Choosing core vs advanced usage

• Core path (recommended): use scrape/crawl with minimal options (runtime, conversion.output_format, traversal.mode="list"). This keeps behavior predictable and low-risk for most agent workflows.
• Advanced path (explicit opt-in): use deep traversal, custom dispatcher controls, JS transforms, extraction schemas, and artifact capture only when required by task outcomes.
• Safety budgets and gates: inspect config://server for settings.defaults, settings.limits, settings.policies, and settings.capabilities to understand active constraints and feature gates before enabling advanced options.

Resources

Prompts

Architecture

graph TD
    A[MCP Client] -->|stdio / HTTP| B[FastMCP v3]
    B --> C[Tool Router]
    C --> D[scrape]
    C --> E[crawl]
    C --> F[close_session]
    C --> G[get_artifact]
    D & E & F & G --> N[AsyncWebCrawler Singleton]
    N --> O[Headless Chromium]
    B --> P[Resources]
    B --> Q[Prompts]

    style B fill:#4B8BBE,color:#fff
    style N fill:#FF6B35,color:#fff
    style O fill:#2496ED,color:#fff

The server uses a single-module architecture:

• FastMCP v3 handles MCP protocol negotiation, transport, tool/resource/prompt registration, and message routing
• Lifespan-managed AsyncWebCrawler starts a headless Chromium browser once at server startup and shares it across all tool invocations, then shuts it down cleanly on exit
• 4 tool functions decorated with @mcp.tool() define the canonical surface
• 2 resource functions decorated with @mcp.resource() return JSON
• 3 prompt functions decorated with @mcp.prompt return structured Message lists

[!IMPORTANT] There are no intermediate manager classes or custom HTTP clients. The server delegates all crawling to crawl4ai's AsyncWebCrawler and all protocol handling to FastMCP. Only 2 runtime dependencies.

Configuration

Testing

# Run all tests
uv run pytest

# Coverage gate (>=90%)
uv run pytest --cov=crawl4ai_mcp --cov-report=term-missing

# Smoke tests only
uv run pytest -m smoke

# Unit tests only
uv run pytest -m unit

# Integration workflow tests
uv run pytest -m integration

# End-to-end workflow tests
uv run pytest -m e2e

# Manual live test (requires browser)
uv run python tests/manual/test_live.py

[!NOTE] All automated tests run in-memory using fastmcp.Client(mcp) — no browser or network required. The test suite mocks AsyncWebCrawler for fast, deterministic execution.

Contributing

See the Contributing Guide for details on setting up your development environment, coding standards, and the pull request process.

License

This project is licensed under the MIT License. See the LICENSE file for details.