Firecrawl MCP

What it does

Claude alone cannot read the live web past its training cutoff. The usual workaround is to spin up a scraper or headless browser, post-process the result, and paste it back into context — each site’s JavaScript rendering, markup quirks, and auth flow making that a per-target project.

Firecrawl MCP delegates the whole step. Claude calls the tools directly, and Firecrawl returns clean, LLM-ready markdown or structured JSON ready to drop into context.

Key features — 12+ tools

• Scraping — firecrawl_scrape / firecrawl_batch_scrape / firecrawl_check_batch_status

  Extract a single URL into JSON, markdown, or branding format, or process many known URLs in parallel. Batch runs asynchronously; poll progress with firecrawl_check_batch_status.

• Site discovery — firecrawl_map / firecrawl_crawl / firecrawl_check_crawl_status

  Use firecrawl_map to discover indexed URLs on a site, and firecrawl_crawl to walk multiple pages with depth and page limits. Crawl runs asynchronously and is polled.

• Web search — firecrawl_search

  Returns search results with optional content extraction for each hit.

• Structured extraction — firecrawl_extract

  Applies an LLM against the target page to produce JSON conforming to a schema you provide — useful when text-only output isn’t enough.

• Autonomous research agent — firecrawl_agent / firecrawl_agent_status

  Hand the agent a topic and it runs multi-source research asynchronously. Poll progress with firecrawl_agent_status.

• Interactive page automation — firecrawl_interact

  Click, type, and navigate to reach content that static scraping cannot.

• Cloud browser sessions — firecrawl_browser_* (deprecated)

  firecrawl_browser_create, execute, list, delete exist but are marked deprecated. New integrations should prefer the other tools.

Installation

Claude Code (recommended)

claude mcp add firecrawl \
  --env FIRECRAWL_API_KEY="fc-YOUR_API_KEY" \
  -- npx -y firecrawl-mcp

Direct npx

env FIRECRAWL_API_KEY=fc-YOUR_API_KEY npx -y firecrawl-mcp

Claude Desktop · Cursor · Windsurf · VS Code — add to the client’s mcpServers JSON config:

{
  "mcpServers": {
    "mcp-server-firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": { "FIRECRAWL_API_KEY": "YOUR_API_KEY_HERE" }
    }
  }
}

Config locations differ per client — Claude Desktop uses claude_desktop_config.json, Cursor v0.48.6+ via Settings → Features → MCP Servers, Windsurf at ./codeium/windsurf/model_config.json.

HTTP Streamable mode — for shared or remote environments.

env HTTP_STREAMABLE_SERVER=true FIRECRAWL_API_KEY=fc-YOUR_API_KEY npx -y firecrawl-mcp
# Accessible at http://localhost:3000/mcp

Self-hosted Firecrawl — set FIRECRAWL_API_URL to your own Firecrawl instance and the cloud API key requirement is bypassed (unless your instance itself requires authentication).

Environment variables

Variable	Required	Default	Purpose
FIRECRAWL_API_KEY	✅ (for cloud)	—	Firecrawl cloud API key. Generate at https://www.firecrawl.dev/app/api-keys
FIRECRAWL_API_URL		cloud default	Self-hosted instance endpoint — bypasses cloud API when set
FIRECRAWL_RETRY_MAX_ATTEMPTS		3	Retry attempts
FIRECRAWL_RETRY_INITIAL_DELAY		1000ms	Initial retry delay
FIRECRAWL_RETRY_MAX_DELAY		10000ms	Max retry delay
FIRECRAWL_RETRY_BACKOFF_FACTOR		2	Exponential backoff factor
FIRECRAWL_CREDIT_WARNING_THRESHOLD		1000	Warn when remaining credits drop below
FIRECRAWL_CREDIT_CRITICAL_THRESHOLD		100	Critical warning threshold

Notes

• Officially operated by Firecrawl — maintained in the github.com/firecrawl org, distinct from community wrappers.
• Firecrawl is a separate service — a Y Combinator-backed company at firecrawl.dev. Free plan: 1,000 credits/month. Search results and page scrapes cost 1 credit each, interactions 5 credits each. Failed requests typically aren’t charged.
• Self-host option — for full data and cost control, set FIRECRAWL_API_URL to your own Firecrawl instance.
• Built-in retry + credit monitoring — exponential-backoff retries and remaining-credit warnings are tuned via the env vars above.
• Built on the open-source Firecrawl engine — the API this MCP calls is backed by firecrawl/firecrawl, published under AGPL-3.0 (★ 118k). Point it at a self-hosted instance with FIRECRAWL_API_URL to use the same toolset without the cloud.
• A separate plugin exists at claude.com/plugins/firecrawl — same Firecrawl service, different distribution channel: Claude Code-specific slash commands (/firecrawl:scrape, etc.). This MCP server uses standard MCP tool calls (firecrawl_scrape, etc.), which work across Cursor, Windsurf, VS Code, and any other MCP client.
• MIT licensed — free for commercial and personal use.