Firecrawl MCP Server

What it does (plain English)

Firecrawl turns websites into text Claude can work with. Its MCP server (a plug-in that gives Claude a new ability) can scrape one page into Markdown or JSON, map all the URLs on a site, crawl a whole section, search the web, click and type on pages that need interaction, and run an “agent” that researches across several sites and returns structured data. It can also watch a page on a schedule and report changes.

The full server lists 27 tools by default. The ones you’ll use most:

Tool	Use it when	Returns
firecrawl_scrape	You know the exact URL	JSON or Markdown
firecrawl_map	You need the list of pages on a site	List of URLs
firecrawl_crawl	You want content from many pages under a site	Crawl results
firecrawl_search	You start from a question, not a URL	Ranked results
firecrawl_interact	The page needs a click or typing first	Result plus a scrape ID
firecrawl_agent	The answer spans several unknown sites	Structured JSON

There are also specialised searches for programming questions, US government and legal material, and scientific papers.

Is it for you?

Pick it if you collect information from the web: competitor pricing pages, product listings, documentation, research papers. It’s stronger than Claude’s built-in web fetch when you need many pages, structured fields, or pages that only load after JavaScript.

You can try three tools (scrape, search and parse) with no account through the hosted keyless tier. For crawling, mapping or the research agent you need a Firecrawl account, and usage draws on Firecrawl credits.

What you can do with it

• Structured extraction: “Scrape these five competitor pricing pages and give me a table of plan name, price and seat limit.” Use JSON format with a schema for clean fields.
• Site inventory: “Map example.com and list every URL under /blog.”
• Bounded crawl: “Crawl the docs section, max 50 pages, and summarise what each page covers.” Set limit and includePaths so it doesn’t run away.
• Research: “Find three recent papers on sleep and memory and summarise their methods” via the firecrawl_research_* tools.
• Change tracking: ask Claude to create a monitor with firecrawl_monitor_create on a page and get diffs when it changes.

How to set it up

1. claude.ai. Install the native connector from https://claude.ai/directory/connectors/firecrawl and sign in to Firecrawl. You’ll choose which Firecrawl team to approve.

2. Claude Code (sign in):

   claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp-oauth

   Then complete sign-in through /mcp. The OAuth URL is a server address for your client, not a page to open in your browser.

3. API key instead. Keep the server URL as https://mcp.firecrawl.dev/v2/mcp and set the header Authorization: Bearer <FIRECRAWL_API_KEY> in your client’s settings. Firecrawl says never to put the key in the URL or in a chat message.

4. Claude Desktop, local server. Add this to claude_desktop_config.json (needs Node.js), then restart:

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

5. Manage access. Review and revoke connections at https://www.firecrawl.dev/app/settings?tab=mcp.

Watch out for

• Credits. Crawls and agent runs can use many credits. Always bound crawls with limit, include/exclude paths or maxDiscoveryDepth, and ask Claude to check firecrawl_credit_usage.
• Scraped text is untrusted. Web pages can contain instructions aimed at AI. Treat extracted content as data, and keep approvals on for any tools that write elsewhere.
• Keyless is limited. On the keyless endpoint only scrape, search and parse are available, and they’re rate-limited.
• Interactive sessions are one turn at a time. Each firecrawl_interact call runs one prompt or code turn and returns; it isn’t a step-by-step remote control like Playwright.
• Self-hosting. You can point the local server at your own Firecrawl instance with FIRECRAWL_API_URL.

Alternatives

• Exa MCP for search-first research with clean page fetches.
• Playwright MCP when you need to drive a browser through a flow, not just read pages.
• Cloudflare’s Browser Run server for fetching pages and screenshots from Cloudflare’s infrastructure.