Firecrawl MCP server

The Firecrawl MCP server gives AI agents live web access through Firecrawl: search the web, scrape a URL into Markdown or JSON, map and crawl whole sites, and click through pages.

Checked by CookiesForAI on : repository facts (stars, last push, archived) re-read from GitHub; listing in the MCP Registry checked.

Written by

Connect Firecrawl

Claude Code, signing in with your Firecrawl account (finish sign-in with /mcp)

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

Codex, signing in with your Firecrawl account

Terminal
codex mcp add firecrawl --url https://mcp.firecrawl.dev/v2/mcp-oauth

Cursor or any client with a JSON config, keyless

Config
{"mcpServers": {"firecrawl": {"url": "https://mcp.firecrawl.dev/v2/mcp"}}}

Run the open-source server locally with an API key

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

In ChatGPT and Claude.ai, Firecrawl points to its native connector in each app's connector directory instead of a URL. With an API key on the hosted endpoint, send it as an Authorization: Bearer header in the client config, never in the URL.

Or paste this into your coding agent: Connect the Firecrawl MCP server to my coding agent using its documented setup: https://github.com/firecrawl/firecrawl-mcp-server. Checked by CookiesForAI on : repository facts (stars, last push, archived) re-read from GitHub; listing in the MCP Registry checked.

Your first question with this server installed

5,542tokens sent to the model with your first question: the tool list plus one tool answer

FirecrawlTool list 3,920One answer 1,6223 tools listed without a key

Typical serverFirst question 2,022 (median of 10 servers)

Firecrawl's first question takes 5,542 tokens: its tool list plus one answer. That is about 2.7 times the median first question of the 10 servers we measured with both numbers on (2,022 tokens).

Measured without a key or sign-in. Assistants that load every tool send this list to the model with each question; some load tools only when a task needs them. The one answer is from firecrawl_search, a fixed read-only call. Our page names 12 of its tools, so set up as its maker documents, it can list more than we measured. Tokens are counted with OpenAI's o200k tokenizer; other models count a little differently.

Measured . How we measure.

What the Firecrawl MCP server does

The Firecrawl MCP server connects an AI agent to Firecrawl, a web data API. With it the agent can run a web search and get ranked results, scrape a known URL into clean Markdown or into JSON that matches a schema you give it, list every URL on a site without fetching them, and crawl a section of a site within limits you set. It also parses files such as PDFs into text.

Firecrawl's server goes past plain fetching. An interact tool clicks, types and navigates on a page before reading it, an agent tool runs multi-source research and returns structured data when you do not know which sites hold the answer, and monitor tools re-check a page on a schedule and report what changed. A developer search tool looks across GitHub issues, pull requests, READMEs and docs for programming questions.

Firecrawl runs the server for you at mcp.firecrawl.dev, so most people connect a URL instead of installing anything. The same open-source code runs locally with npx over stdio or Streamable HTTP, and it can point at a self-hosted Firecrawl instance. The repo had 7,538 GitHub stars on .

Tools it gives your agent

  • firecrawl_scrape
  • firecrawl_search
  • firecrawl_map
  • firecrawl_crawl
  • firecrawl_check_crawl_status
  • firecrawl_parse
  • firecrawl_agent
  • firecrawl_agent_status
  • firecrawl_interact
  • firecrawl_interact_stop
  • firecrawl_developer_search
  • firecrawl_credit_usage

We also list this company's other pages: Firecrawl skill.

When to use Firecrawl

  • Your agent needs the current text of a web page, not what the model remembers.
  • You want product prices, listings or contacts pulled from a site into JSON.
  • You need every docs page of a site crawled and fed to the agent as Markdown.
  • A page only shows its content after a click or a login step.
  • You want to try web scraping in an agent with no account at all, through the keyless endpoint.

When to pick something else

  • Heavy use: crawl, map and agent need an API key or sign-in, and every call spends credits from a monthly allowance.
  • Driving a browser step by step across a long session: each interact call runs one turn and returns. Playwright or Chrome DevTools fit that job better.

What Firecrawl needs

  • A Firecrawl account and API key for the full tool set (keyless covers search, scrape and parse)
  • Node.js, only when running the server locally with npx

What the Firecrawl MCP server costs

Free and open source under MIT, with a keyless hosted tier that has daily limits; full use spends Firecrawl credits from a free or paid plan.

Current prices: www.firecrawl.dev . We do not list prices because vendors change them often. Check the publisher's own page.

Questions people ask

Is the Firecrawl MCP server free?

The server code is free under MIT, and the hosted keyless endpoint lets an agent search, scrape and parse without an account, with daily limits. Beyond that you spend Firecrawl credits from the monthly allowance of the Free plan or a paid plan (pricing page, ).

Does the Firecrawl MCP server work with ChatGPT?

Yes. Firecrawl's docs send ChatGPT users to a native Firecrawl connector in ChatGPT, and give setup lines for Codex, Claude Code, Cursor, OpenCode and any client that reads a JSON MCP config.

Can I use it with a self-hosted Firecrawl?

Yes. Run the server locally and set FIRECRAWL_API_URL to your own instance; the API key then becomes optional.