By the fastCRW team · Tutorial · Tested on 2026-11-20 with fastCRW 0.46.3 and no Node.js on PATH.
If you want Firecrawl's MCP tools on your own machine, you need Node.js. Firecrawl's local MCP docs list Node.js 22 or newer, and the new Firecrawl Local binary does not change that: its README says firecrawl-local mcp starts the local API and runs the firecrawl-mcp npm package through npx. That is fine on a developer laptop. It is a problem in a slim CI image, a locked-down corporate machine, or a server where you do not want a JavaScript runtime just to fetch web pages. fastCRW's MCP server is compiled into its Rust binary, so the only thing to install is that binary.
What each option needs
| Setup | Local runtime | Where scraping runs |
|---|---|---|
| Firecrawl hosted MCP URL | None | Firecrawl's servers |
npx -y firecrawl-mcp | Node.js 22+ | Firecrawl Cloud, or a Firecrawl server you point it at |
firecrawl-local mcp | Node.js 22+ (plus Rust and a C compiler to build the binary) | Your machine for scrape, crawl, map and batch scrape; Firecrawl Cloud for search and blocked pages |
crw mcp (fastCRW) | None beyond the binary | Your machine, unless you add a fastCRW Cloud key |
Install and register the server
curl -fsSL https://fastcrw.com/install | sh
which crw
Add crw mcp as a stdio server in your MCP client, using the full path from which crw:
{
"mcpServers": {
"crw": {
"command": "/usr/local/bin/crw",
"args": ["mcp"]
}
}
}
If you would rather not edit config files, crw setup registers the server with Claude Code, Cursor, Codex, Gemini CLI, OpenCode and Windsurf when they are already installed. Restart the client afterwards so it loads the new server.
Proof: an MCP session with no Node.js
To make sure nothing quietly shells out to npx, we started the server with an empty environment and a PATH of only /usr/bin:/bin, where command -v node finds nothing. Then we sent three JSON-RPC messages over stdin: initialize, notifications/initialized and a tools/call for crw_scrape:
env -i HOME="$HOME" PATH=/usr/bin:/bin /usr/local/bin/crw mcp
The server answered initialize as {"name":"crw-mcp","version":"0.46.3"}, and the crw_scrape call returned:
{"jsonrpc":"2.0","id":3,"result":{"content":[{"text":"{\"creditCost\":1,\"markdown\":\"# Example Domain\\n\\nThis domain is for use in documentation examples without needing permission. ...
A separate tools/list call on the same plain install listed crw_scrape, crw_crawl, crw_check_crawl_status, crw_map, crw_extract, crw_check_extract_status, crw_cancel_extract, crw_parse_file and crw_enrich. crw_search joins the list once you configure a search backend or a fastCRW Cloud key. The three extract tools need an LLM: without one, crw_extract answers that extraction requires an LLM set in [extraction.llm].
Keep the Firecrawl API too
MCP is not the only way agents call a scraper. The same binary serves the Firecrawl v2 API, so scripts written for Firecrawl or Firecrawl Local keep working:
crw serve --host 127.0.0.1 --port 3002
That is the whole migration from Firecrawl Local for SDK users. The step-by-step version, with the Python, JavaScript and curl runs, is in Move from Firecrawl Local to fastCRW in one line, and the feature-by-feature table is in Firecrawl Local vs fastCRW.
Prefer a hosted MCP endpoint?
fastCRW Cloud also runs a managed MCP endpoint at https://fastcrw.com/mcp, with OAuth sign-in for clients that support it. It needs no local runtime at all, and new accounts start with 1000 free credits.
