Firecrawl MCP Server: Web Scraping for AI Agents

The official Firecrawl listing points to firecrawl/firecrawl-mcp-server. It connects an MCP client to Firecrawl's web extraction service. Use the maintainer documentation to choose a connection before requesting a crawl.

Docs-derived; sources checked October 17, 2026. MCPFind did not authenticate, scrape a page, or measure credit consumption. The listing's May be outdated warning remains relevant: check its current package and setup against the official repository.

Choose hosted access or local stdio

The official repository documents hosted OAuth/API-key connection options and a local stdio package. Use the hosted instructions for a client that supports that route, following its sign-in or credential configuration. Do not paste a local command configuration into a remote URL field.

For a local stdio client, the documented package is firecrawl-mcp with FIRECRAWL_API_KEY:

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

Adapt this example to your client's documented configuration format and supply the key through its credential mechanism. Keep real credentials out of shared files. Local stdio launches a client-side process; it does not make the hosted extraction service local. Review current account limits, billing, and the available tools in the official setup documentation.

Verify one public page before a crawl

Use a public documentation site you are allowed to access, a known topic, and a one-page extraction goal. For example: find the installation page and return only its title, source URL, and documented install command.

1. Ask firecrawl_map for candidate URLs on the named site. Inspect the results and choose one public documentation URL; do not expand into an unattended site-wide crawl.
2. Ask firecrawl_scrape for that URL with output limited to the fields required for the question. Use the tool's current schema to select the supported output format and extraction options.
3. Expect a title, the selected source URL, and the installation text or structured fields requested. Open the publisher page and compare the command and surrounding requirements with the extraction.
4. Inspect any returned usage information and the Firecrawl dashboard before expanding the task. If extraction is empty or blocked, record that outcome instead of asking the agent to fill the gap from memory.

A map identifies candidate links; a scrape extracts one selected page; a crawl follows multiple pages and needs a separate scope and usage decision. Tool availability and usage fields can change, so inspect the connected tool schema rather than relying on a fixed tool count.

Review the next operation separately

This first task reads a public page. Expanding to many pages, metered research, interactive browsing, authenticated content, or scheduled monitoring changes its scope. Review the domains, page limit, account usage, and any form or write action before authorizing that work. A successful scrape does not establish permission to access other content or guarantee that extracted text is complete.

Browse devtools integrations for alternatives, then compare the same public page, required fields, and usage limits. For protocol basics, start with what MCP is.