Firecrawl MCP Server

What firecrawl/firecrawl-mcp-server MCP server does

The firecrawl/firecrawl-mcp-server MCP server gives MCP-compatible agents access to Firecrawl's web retrieval and automation functions. It can search general web, news, image, research, GitHub, and developer-oriented indexes; fetch a known URL; enumerate URLs within a site; and collect content from multiple pages. Scraping can return markdown, HTML, links, screenshots, branding information, targeted answers, or JSON shaped by a supplied schema.

The server also supports document parsing for PDFs, Word files, RTF, OpenDocument files, spreadsheets, and common HTML inputs. Research tools search paper metadata, inspect canonical paper records, find related papers, and retrieve relevant passages when indexed full text is available. For longer-running work, an asynchronous research agent can search, navigate, read pages, and assemble a structured response after its job is polled to completion.

How it works

Known-page requests generally use firecrawl_scrape, while firecrawl_search discovers sources and firecrawl_map lists URLs without retrieving every page body. firecrawl_crawl starts a bounded multi-page job and firecrawl_check_crawl_status reads its progress and results. Crawl scope can be constrained by paths, depth, page count, sitemap behavior, subdomains, external links, and delays.

The server can reuse recently indexed content. A scrape therefore does not necessarily represent the page's current state; maxAge: 0 requests a live fetch. Search-attached scraping follows a separate reuse behavior, so use the direct scrape tool when freshness matters. Live interaction uses a browser session to navigate, click, fill forms, or run bounded Bash, Python, or Node code. Such actions can change external systems, including through form submission.

Setup and configuration

For local stdio execution, the documented package command is npx -y firecrawl-mcp. Cloud use normally requires FIRECRAWL_API_KEY; FIRECRAWL_API_URL can point to a self-hosted Firecrawl instance instead. A static OAuth access token can be supplied through FIRECRAWL_OAUTH_TOKEN for stdio operation. The project also documents local Streamable HTTP mode, enabled with HTTP_STREAMABLE_SERVER=true, at http://localhost:3000/mcp.

The hosted MCP endpoint is https://mcp.firecrawl.dev/v2/mcp. Its keyless free tier allows scrape, search, and parse with rate limits, while tools such as crawl, map, and agent require a key. OAuth clients can use the separate connection endpoint at https://mcp.firecrawl.dev/v2/mcp-oauth; API-key clients send a Bearer authorization header rather than putting the key in the URL.

Tools and capabilities

The firecrawl/firecrawl-mcp-server MCP server includes tools for:

• Searching web, news, images, developer sources, GitHub content, and indexed research papers.
• Scraping known URLs and parsing uploaded or local documents into selected formats.
• Mapping site URLs and crawling bounded collections of pages.
• Running asynchronous research jobs and checking their status.
• Opening live browser sessions, performing interactions, and stopping those sessions.
• Creating, listing, updating, running, inspecting, and deleting recurring monitors.
• Comparing monitor checks with retained predecessors using markdown diffs, JSON field changes, or mixed tracking.
• Inspecting papers, finding related citation-graph candidates, and retrieving relevant paper passages.

Limitations and notes

The deprecated firecrawl_extract entry point should be replaced with firecrawl_scrape using JSON output and schema options. Agent research is asynchronous and initially returns only a job ID; callers must poll its status, typically waiting 15–30 seconds between checks. Crawl responses can be large, so conservative page and depth limits are appropriate when full-site coverage is unnecessary.

Some hosted and authenticated features have different requirements. Zero-data-retention parsing requires an eligible authenticated account, and hosted document parsing uses a two-step upload flow. Safe mode does not support webhooks or interactive scrape actions. Monitor deletion is permanent, and live browser actions may produce persistent side effects.