apps/local of the Firecrawl repository.
Choose Firecrawl Local, self-hosting or Firecrawl Cloud
Install
You need Rust and a C compiler. On macOS, install the Xcode command line tools withxcode-select --install. On Debian and Ubuntu, install build-essential. Then install Firecrawl Local from crates.io.
apps/local. The binary lands in target/release/firecrawl-local.
- Chrome or Chromium. Firecrawl Local renders pages that need JavaScript in the browser already installed on your machine. Without one, static pages still work and a page that needed a browser comes back with a warning saying so.
- Node.js 22 or newer. The MCP server is the
firecrawl-mcpnpm package, which Firecrawl Local runs throughnpx.
Quick start
Scrape a page and print its markdown.--markdown, scrape prints the same JSON the Firecrawl API returns. Ask for more formats with --format markdown,links. The choices are markdown, html, rawHtml and links. PDFs and Office documents are parsed on your machine too.
Add it to your AI agent
init adds Firecrawl Local as an MCP server to every AI tool it finds on your machine.
--dry-run to see the changes without writing them, or --agent cursor to set up one tool. Restart the tool afterwards so it loads the new server.
To add it by hand, register firecrawl-local mcp as a stdio MCP server. In a JSON config that looks like this, with the full path to the binary.
Run the Firecrawl API locally
serve answers the Firecrawl v2 API on port 3002.
http://localhost:3002. No API key is needed for local routes.
Every other
/v2 route, such as /v2/search, is passed on to Firecrawl Cloud unless FIRECRAWL_CLOUD says otherwise. The v0 and v1 routes are not served. Crawl and batch jobs live in memory and expire an hour after they start.
Commands
Run
firecrawl-local <command> --help for every option.
What runs where
Scraping, crawling, mapping and document parsing run on your machine and send nothing to Firecrawl. Three things need Firecrawl Cloud because a laptop does not have the index, the network or the data.
A page Firecrawl Cloud fetched for you carries
"source": "firecrawl-cloud". If the cloud cannot get it either, the result has "blocked": true and a warning that says why. Cloud calls carry a random install ID and the name of the tool that made them, which is how Firecrawl counts installs.
To sign in, run login. It opens the browser and saves your API key in the same place the Firecrawl CLI keeps it, so either tool’s login works for both.
serve picks up the login straight away, and so do the MCP server’s search and scrape tools. The MCP server’s Alexandria tools check for a key when they start, so restart your AI tool after signing in to use them.
FIRECRAWL_CLOUD controls the hand-off.
An unrecognised value counts as
off. login always talks to Firecrawl, because signing in is its job.
Security
The defaults differ between the tools you run yourself and the API server, because the API takes requests from other programs.scrape,crawl,mapandmcprun for the person at the keyboard, so they can reachlocalhostand your private network.scrape,crawlandmcptake--public-onlyto refuse those addresses.servelistens on127.0.0.1and refuses localhost and private addresses unless you pass--allow-private. Binding another address with--hostrequires--token, and requests are checked against an allowed Host list that--allowed-hostadds to.- Always refused. Link-local and cloud metadata addresses are blocked in every mode, including after redirects and DNS resolution.
robots.txt when crawling, where site owners can address it as FirecrawlLocal.
Running in Docker
Chrome keeps its sandbox on, which needs user namespaces, and Docker’s default seccomp profile blocks them. To render JavaScript pages in a container, use a profile that keeps Docker’s defaults and also allowsclone, setns and unshare. Playwright publishes one.
seccomp=unconfined for this. It removes the whole filter from a container that loads pages you do not control.
Configuration
Limits
- Local scrapes skip the
jsonandsummaryformats with a warning, and ignore browseractions. Screenshots need Chrome. - Crawl and batch jobs are kept in memory, so they end when
servestops. - The v0 and v1 API routes are not served.

