Use Firecrawl MCP for Web Scraping & Search in Cursor + Claude

Set up the Firecrawl MCP server in Cursor to give Claude real-time web scraping and search tools — no manual browsing needed during AI-assisted coding.

The Firecrawl MCP server turns Claude inside Cursor into a research-capable coding partner. Instead of pasting documentation snippets manually, Claude can fetch live pages, crawl entire sites, and run web searches through structured tool calls — all from inside Agent mode. This guide walks through the full wiring, from API key to a working test prompt. If you are new to MCP in general, start with our Cursor MCP setup guide to understand how the protocol connects tools to your editor.

Sign up at firecrawl.dev and open your dashboard to generate an API key. The free tier includes enough credits to experiment with scraping and search endpoints while you evaluate the integration. Copy the key somewhere safe — treat it like any other secret, because it authorises billable requests against your Firecrawl account.

# Store it as an environment variable (optional but recommended)
export FIRECRAWL_API_KEY="fc-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

The Firecrawl MCP server ships as an npm package that runs over stdio. You do not need to install it globally — npx will download and cache the latest version the first time Cursor spawns it. Running a quick version check now confirms your Node.js toolchain can fetch and execute the package before you introduce it into Cursor's config.

# Verify it runs correctly before wiring it into Cursor
npx firecrawl-mcp --version

In Cursor, open the Command Palette with Cmd+Shift+P (macOS) or Ctrl+Shift+P (Windows/Linux) and search for "Open MCP Settings". This opens the mcp.json configuration file where all MCP servers are registered. If the file does not exist yet, Cursor will create it for you in your user settings directory.

Add a new entry inside the mcpServers object in mcp.json. The command tells Cursor how to spawn the process, and the env block passes your API key without hardcoding it into any tracked file. Use the -y flag on npx to auto-confirm the package install prompt.

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

Save mcp.json and reload the Cursor window via Cmd+Shift+P → Reload Window. Reopen MCP Settings and look for a green connected status next to the firecrawl entry. If it shows red, expand the log — the most common issues are a missing API key, an outdated Node.js version, or a network policy blocking npx downloads.

Open a Cursor chat panel in Agent mode with Claude selected as the model. Ask it to scrape a URL, run a web search, or crawl a documentation site. Claude will detect the Firecrawl tools automatically and invoke scrape, search, or crawl depending on the intent — no manual tool selection required.

// Example prompts to test Firecrawl tools via Claude
// "Scrape https://example.com and summarise the main content"
// "Search the web for the latest Next.js 15 release notes"
// "Crawl https://docs.example.com and extract all API endpoint paths"

Once Firecrawl is wired up, your Claude-powered features depend on the model actually selecting the right tool for each prompt. Regressions here are silent: the model might answer from stale knowledge instead of scraping fresh data, and users only notice when the output is wrong. desplega.ai lets you write natural-language E2E tests that assert which MCP tool was called, what arguments were passed, and whether the response shape matches expectations — so a bad prompt change or a broken MCP config fails CI instead of production.

If you are combining Firecrawl with other servers, review our notes on MCP best practices to keep tool selection deterministic and your agent behaviour reproducible across runs.