Firecrawl just killed the Docker stack for local scraping. On November 18, the team launched Firecrawl Local – a single Rust binary that replaces the entire five-service Docker Compose setup and serves the same v2 API on localhost:3002. It idles at 11 MB of memory. That is not a typo.
What the Old Setup Cost You
If you have tried self-hosting Firecrawl before, you know the tax. The Docker Compose stack required five services running at once: the Node.js API, a Playwright service for JavaScript rendering, Redis, RabbitMQ, and Postgres. No documented upgrade path. Around 2.5 GB of RAM at idle, before any real work started.
Most developers either paid for the cloud API or skipped local deployment entirely. Neither is great when you want scraping in your dev loop or wired into an AI agent that handles sensitive data.
What Firecrawl Local Actually Is
Firecrawl is the web-to-markdown API that ships clean data to LLMs. It handles JavaScript rendering, proxy rotation, and anti-bot bypass so you do not have to. The core API covers scraping a single page, crawling a full domain, and mapping all URLs on a site. Over 1.25 million developers use the hosted version.
Firecrawl Local is a Rust rewrite of the scraping primitives that ships as one binary. No Docker. No database. No queue. Install it with Cargo, run it, and it answers the same v2 API on localhost:3002.
cargo install firecrawl-local
firecrawl-local
The performance claims are significant: 37x faster scraping than the old self-hosted version, and 230x less memory at idle. The team has not published benchmark methodology or hardware specs, so treat those numbers as directionally true rather than gospel. What is not in question: 11 MB idle versus a 2.5 GB Docker stack is a material difference regardless of the exact multiple.
MCP Integration Takes One Command
This is the part that matters most if you are building with Claude Code or Cursor. Run:
firecrawl-local init
That registers the binary as an MCP server in both tools. No editing JSON config files, no manual path setup. Claude Code and Cursor get live web access immediately. The MCP server routes scrape, crawl, and map calls locally; blocked pages and search fall back to Firecrawl Cloud if you have an API key set.
If you do not want anything leaving your machine, set FIRECRAWL_CLOUD=off. That disables search, the blocked-page fallback, and Alexandria – but scrape, crawl, and map stay fully local with no API key required.
What You Do Not Get
Firecrawl Local is not a full replacement for the self-hosted version. It does not support structured extraction via JSON schema, site monitors, or the Alexandria data layer. JavaScript-heavy pages still require Chrome or Chromium installed on the machine. The binary does not bundle a browser. The MCP server also depends on Node.js 22 or newer.
The full self-hosted Docker Compose stack remains the right choice for production crawl jobs, schema-based extraction, or anything that needs Alexandria.
The Right Tool for Each Job
Here is where Firecrawl Local fits:
- Local development: Drop it into your dev environment with zero overhead
- AI agent workflows: Wire it into Claude Code or Cursor via firecrawl-local init in 60 seconds
- Privacy-sensitive scraping: Run with FIRECRAWL_CLOUD=off for fully local operation
- Lightweight CI: 11 MB idle, no Docker required, runs anywhere Rust binaries run
If you need structured JSON extraction from schemas, production-scale crawls, or the Alexandria provider network, the full self-hosted version or the cloud API is still the answer.
The PR that introduced Firecrawl Local (firecrawl/firecrawl #5685) landed on November 18. The quickstart docs walk through installation and MCP setup. The Hacker News thread has community caveats if you want the unfiltered take before you switch.

