Firecrawl MCP Server

What it does

This is Firecrawl's own MCP server for its web data API. An assistant can search the web, turn a known URL into markdown or into JSON that fits a schema you pass, list a site's URLs, crawl a bounded set of pages, and click or type on a page before reading it. It also parses PDFs, Word files and spreadsheets.

Beyond the open web there are three hosted indexes: one for coding questions (GitHub issues, merged pull requests, READMEs and docs), one for US statutes, regulations and court opinions, and one for papers from PubMed, bioRxiv, medRxiv and arXiv, with tools to read a paper and walk its citations. Other tools start a research agent job, manage scheduled page monitors that send diffs by webhook or email, browse the Alexandria catalogue of data providers, and show credit usage.

You can connect to Firecrawl's hosted endpoint, or run the npm package locally over stdio. Locally it talks to Firecrawl's cloud with your key, or to your own Firecrawl instance through FIRECRAWL_API_URL. Started with neither a key nor a URL, it runs keyless: src/index.ts logs that only firecrawl_scrape and firecrawl_search work then, rate-limited per IP. The README counts 27 tools, but the repository's stdio contract snapshot lists 28 with an API key set. At start the server also loads a .env file from its working directory if one is there, which the README does not mention.

What it needs

Create an API key at https://www.firecrawl.dev/app/api-keys and set it as FIRECRAWL_API_KEY. For a self-hosted Firecrawl, set FIRECRAWL_API_URL instead (a key only if your instance asks for one). On the hosted endpoint, sign in through the client with https://mcp.firecrawl.dev/v2/mcp-oauth, or send Authorization: Bearer <FIRECRAWL_API_KEY> to https://mcp.firecrawl.dev/v2/mcp.

Limits

• Limit Run locally without FIRECRAWL_API_KEY or FIRECRAWL_API_URL, it starts keyless: firecrawl_scrape and firecrawl_search work, rate-limited per IP, and other tools answer Unauthorized.
• Not covered The hosted endpoint without a key exposes only firecrawl_scrape, firecrawl_search and firecrawl_parse. Crawl, map, agent and the rest need a key or OAuth.
• Cost Calls spend Firecrawl credits. firecrawl_search with scrapeOptions bills every fetched result as a separate scrape.
• Limit In local HTTP mode, requests whose Host or Origin header is not localhost, 127.0.0.1, [::1], a non-wildcard HOST or a FIRECRAWL_MCP_ALLOWED_HOSTS entry get 403.
• Not covered Each firecrawl_interact call runs one prompt or code turn to the end. A browser session cannot be driven step by step from the client inside one call.
• Not covered Several URLs cannot go in one scrape call. Batch scraping is only on the Firecrawl API outside MCP.
• Limit firecrawl_extract and firecrawl_research_search_github are deprecated and return an error naming firecrawl_scrape, firecrawl_agent or firecrawl_developer_search instead.
• Rate limit Search feedback refunds a credit per search, up to 100 credits per team per UTC day by default.

Who it suits

A good fit if you want web pages as markdown or structured JSON without running a browser yourself, and are fine with URLs and queries going to Firecrawl. The hosted endpoint needs nothing installed and lets you try scrape and search before signing up. The local package is the way to point it at a self-hosted Firecrawl instance. Developers, lawyers and researchers also get the code, government and paper indexes in the same server. Look elsewhere if pages must not leave your network to a third party, or if you need to hold one browser session and drive it step by step with your own retry logic, which the README itself says this server does not do.