generated: '2026-08-29'
method: probed
source: >-
  POST https://gate.crawl4ai.com/mcp {"jsonrpc":"2.0","id":1,"method":"tools/list"}
  -> HTTP 200 text/event-stream, 2026-08-29
provider: Crawl4AI
providerId: crawl4ai
name: Crawl4AI Cloud MCP Server
description: >-
  Crawl4AI ships a hosted, streamable-HTTP MCP server at
  https://gate.crawl4ai.com/mcp that exposes the Cloud API as five native agent
  tools. tools/list answers ANONYMOUSLY — the full tool set with complete JSON
  Schema inputSchemas came back on an unauthenticated POST — while tools/call
  requires a key. The separately-distributed Claude Code plugin also ships a
  local stdio server, so both deployment modes exist.
deployment:
  mode: both
  endpoint: https://gate.crawl4ai.com/mcp
  install: >-
    claude mcp add --transport http crawl4ai https://gate.crawl4ai.com/mcp
    --header "Authorization: Bearer sk_live_..."
  package: https://github.com/unclecode/crawl4ai-cloud-sdk/tree/main/python/claude-plugin
  auth: api-key
  verified: probed
transport:
  remote:
    protocol: streamable-http
    endpoint: https://gate.crawl4ai.com/mcp
    methods_observed:
      - method: GET
        status: 405
        note: Method Not Allowed — POST only
      - method: POST tools/list
        status: 200
        content_type: text/event-stream
        note: Returned all five tools with inputSchema, unauthenticated
    auth:
      style: bearer
      header: 'Authorization: Bearer sk_live_...'
      alternate_header: 'x-api-key: sk_live_...'
      required_for: tools/call
      anonymous_for: tools/list
  local_stdio:
    source: >-
      https://github.com/unclecode/crawl4ai-cloud-sdk/blob/main/python/claude-plugin/.mcp.json
    command: bash
    args:
      - '${CLAUDE_PLUGIN_ROOT}/start.sh'
    note: >-
      Shipped inside the first-party Claude Code plugin (crawl4ai-claude-plugins
      marketplace). Runs in cloud or local mode; /crawl4ai:switch flips between
      them. This is a human-install surface, not an agent-reachable URL.
clients_documented:
  - Claude Code
  - Cursor
  - Windsurf
tools:
  - name: scrape
    description: Scrape a web page and return clean markdown or HTML.
    required: [url]
    parameters: [url, format, proxy, country, parse, bypass_cache]
    rest_equivalent: POST /scrape
  - name: search
    description: >-
      Search the web (browser-free, multi-engine) and return ranked results.
      rich=true adds follow-up questions, related queries and an entity card.
    required: [q]
    parameters: [q, rich]
    rest_equivalent: GET /search
  - name: answer
    description: >-
      EXPERIMENTAL: get a direct answer to a question with its sources; `answered`
      is false when no direct answer exists.
    required: [q]
    parameters: [q]
    rest_equivalent: GET /answer
  - name: extract
    description: >-
      Extract structured data from a page using an LLM. Give a URL plus either an
      instruction or a JSON schema.
    required: [url]
    parameters: [url, instruction, schema, example, bypass_cache]
    rest_equivalent: POST /extract
  - name: batch
    description: Scrape many URLs in one call; returns one result per URL (NDJSON).
    required: [urls]
    parameters: [urls, format, proxy, country, parse]
    rest_equivalent: POST /scrape/batch
tool_count: 5
schemas:
  file: crawl4ai-mcp-tools.json
  note: >-
    Verbatim tools/list JSON-RPC response, including every tool's JSON Schema
    2020-12 inputSchema. Not derived — this is the server's own answer.
divergence:
  note: >-
    The MCP tool set covers only the gate.crawl4ai.com Cloud API. It does NOT
    expose the async job surface (POST /scrape/jobs and friends) and does not
    reach the api.crawl4ai.com v1 API (markdown / screenshot / map / crawl-site)
    at all. See crawl4ai-tool-crosswalk.yml.
self_hosted_mcp:
  note: >-
    The open-source self-hosted Docker server carries its own MCP bridge
    (deploy/docker/mcp_bridge.py) exposing /mcp/sse, /mcp/ws and /mcp/schema on
    the operator's own host and port 11235. It is a templated, customer-run
    endpoint, not a Crawl4AI-operated one, so it is documented but not recorded
    as a reachable agent surface.
  docs: https://docs.crawl4ai.com/core/self-hosting/
maintainers:
  - FN: Kin Lane
    email: kin@apievangelist.com