Firecrawl MCP Server

Overview

Firecrawl MCP Server exposes Firecrawl web data tools through the Model Context Protocol. It supports live web search, single-page and batch scraping, website mapping, crawling, structured extraction, local/self-hosted file parsing, browser/page interaction, autonomous research jobs, monitor management, and research helper tools.

Installation and Transport

Default local package launch uses stdio with npx:

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

The package can also be installed globally with npm install -g firecrawl-mcp. Set HTTP_STREAMABLE_SERVER=true to run the local package over streamable HTTP; current README, npm 3.22.1, and FastMCP source indicate the local endpoint is http://localhost:3000/mcp with default PORT=3000 and HOST=localhost. This is separate from the hosted Firecrawl remote at /v2/mcp; docs that mention local /v2/mcp are not the published package default unless an endpoint override is supplied.

Firecrawl also hosts a streamable HTTP remote at https://mcp.firecrawl.dev/v2/mcp. The remote supports a keyless free tier for scrape/search/interact/parse and optional bearer authentication for full access. Official docs also describe an API-key path form; secret path values are omitted from this registry payload.

Configuration

Use FIRECRAWL_API_KEY for authenticated cloud API access, or FIRECRAWL_OAUTH_TOKEN for a static OAuth access token on stdio. Use FIRECRAWL_API_URL for self-hosted Firecrawl instances. Feedback tools can be disabled with the documented opt-out variables. Retry and credit-warning variables are not included as active package metadata because current npm 3.22.1 runtime and current src/index.ts do not consume them; they only appear in changelog/legacy source evidence.

Capabilities

Reviewed tools include: firecrawl_scrape, firecrawl_batch_scrape, firecrawl_check_batch_status, firecrawl_map, firecrawl_search, firecrawl_search_feedback, firecrawl_feedback, firecrawl_parse, firecrawl_crawl, firecrawl_check_crawl_status, firecrawl_extract, firecrawl_agent, firecrawl_agent_status, firecrawl_interact, firecrawl_interact_stop, firecrawl_monitor_create, firecrawl_monitor_list, firecrawl_monitor_get, firecrawl_monitor_update, firecrawl_monitor_delete, firecrawl_monitor_run, firecrawl_monitor_checks, firecrawl_monitor_check, firecrawl_research_search_papers, firecrawl_research_inspect_paper, firecrawl_research_related_papers, firecrawl_research_read_paper, firecrawl_research_search_github.

Limitations

The keyless hosted remote is rate-limited and does not unlock every tool. Crawl, agent, batch, and monitor workflows can be asynchronous and may require status polling. Large crawls can exceed useful model context. HTTP stream mode requires either credentials or a Firecrawl API URL. Local file parsing needs local/self-hosted execution for direct file access; hosted parse uses upload-reference flow.