Hi, I’m Hoda!
“I want my AI agent to read web pages. But I’m wary of pay-as-you-go APIs, and setting up a server sounds like a hassle.”
For exactly that situation, Firecrawl announced Firecrawl Local on November 18, 2026.
Firecrawl Local is a single binary that runs entirely on your own computer and handles web page scraping, crawling, and mapping (listing a site’s URLs). It needs no database, no queue, and no Docker, and it can connect to AI agents such as Claude Code and Cursor through MCP.
This article covers what Firecrawl Local can and can’t do, how to install and use it, and how it differs from self-hosting and Firecrawl Cloud.
What Is Firecrawl Local?
Firecrawl Local is a standalone Rust binary that lets you run Firecrawl’s scraping features locally.

It fetches pages and extracts just the main content as clean Markdown, all on your own machine. The HTML cleanup uses the same processing as the Firecrawl API.
There are three main ways to use it:
- As a CLI, running
scrape/crawl/mapfrom the terminal - As an MCP server, giving tools to your AI agent
- As a local version of the Firecrawl v2 API, running at
localhost:3002
For the third option, you only need to change the API URL in your existing Firecrawl SDK to http://localhost:3002. Local routes don’t require an API key.
Key Points from the Official Announcement
The official announcement highlights the following:
- Runs Rust-based scrape / crawl / map on your own machine
- 37x faster than the self-hosted version
- 230x lower memory usage (about 11 MB when idle)
- 100% open source
- Works with Firecrawl’s existing MCP server and SDKs as-is
What Runs Locally and What Goes to the Cloud
Firecrawl Local doesn’t run “entirely locally” in every case. Here’s a breakdown of what runs where.
| Feature | Where it runs | Notes |
|---|---|---|
| scrape (single page) | Your PC | HTML / rawHTML / Markdown / links |
| crawl (follow links across a site) | Your PC | Respects robots.txt |
| map (list a site’s URLs) | Your PC | Built from the sitemap and the starting page |
| batch scrape | Your PC | Fetch multiple URLs at once |
| PDF and Office document parsing | Your PC | Parsed locally |
| Search | Firecrawl Cloud | Works without an account, but with usage limits |
| Retrying blocked pages | Firecrawl Cloud | A fallback for sites your PC can’t fetch directly |
| Alexandria (data provider) | Firecrawl Cloud | Requires logging in with a free account |
Pages sent to the cloud through search or blocked-page retries are tagged with "source": "firecrawl-cloud" in the results, so you can always check afterward what left your machine.
Control Cloud Handoff with the FIRECRAWL_CLOUD Environment Variable
You can control what gets passed to the cloud with the FIRECRAWL_CLOUD environment variable.
| Value | Behavior |
|---|---|
auto (default) | Search, blocked-page retries, and Alexandria are all available |
search | Only search goes to the cloud |
off | Nothing is sent to the cloud |
For work you don’t want leaving your machine, such as client projects or internal research, setting it to off is the safest option. Search won’t be available in that case, so you’ll need to supply the URLs you want to read yourself.
How to Install Firecrawl Local
Installation requires Rust and a C compiler. On macOS, run xcode-select --install. On Debian / Ubuntu, install build-essential.
- Set up Rust
Install Rust from rustup. Skip this step if you already have it.
- Install Firecrawl Local
Run the following command in your terminal.
cargo install firecrawl-localThe build takes a few minutes. In one test, the build took just under 6 minutes and the resulting binary was about 19 MB.
- Verify the installation
If a version number is displayed, you’re all set.
firecrawl-local --versionNot required, but these two are handy to have:
- Chrome or Chromium: Lets you fetch pages rendered with JavaScript. Static pages work fine without it
- Node.js 22 or later: Used to run the MCP server (
firecrawl-mcp) vianpx
Basic Usage (CLI)
Fetch a Single Page as Markdown
firecrawl-local scrape https://example.com --markdownWithout --markdown, it outputs the same JSON that the Firecrawl API returns. You can also specify multiple formats, such as --format markdown,links. The available formats are markdown / html / rawHtml / links.
Fetch an Entire Site (crawl)
firecrawl-local crawl https://docs.firecrawl.dev --limit 50 --out ./pagesThis follows links up to 50 pages and saves each page as its own Markdown file in ./pages. The default value for --limit is 100.
List a Site’s URLs (map)
firecrawl-local map https://firecrawl.devThis outputs a list of the URLs the site exposes. It’s handy for getting a sense of how a documentation site is structured.
Use It as a Local API (serve)
firecrawl-local serveThis runs the Firecrawl v2 API at localhost:3002. With the Python SDK, for example, you only need to point it at the new URL:
from firecrawl import Firecrawl
app = Firecrawl(api_url="http://localhost:3002")
print(app.scrape("https://example.com").markdown)These are the API routes available locally:
POST /v2/scrapePOST /v2/mapPOST /v2/crawl(status checks and cancellation supported)POST /v2/batch/scrape(status checks and cancellation supported)GET /health
Other /v2 routes (such as /v2/search) are forwarded to Firecrawl Cloud depending on your settings. v0 and v1 routes are not supported.
Connecting to AI Agents (MCP)
One of the best things about Firecrawl Local is how easy it is to connect to AI agents.
firecrawl-local initRun this command once, and it finds the AI tools installed on your computer and registers itself with them automatically as an MCP server.
These tools are supported:
- Claude Code
- Claude Desktop
- Cursor
- Windsurf
- Gemini CLI
- Codex
Config files are backed up automatically before any changes are made. To preview the changes first, use --dry-run. To set up just one tool, specify it like --agent cursor. After running the command, restart the tool to load the new configuration.
To configure it manually, register firecrawl-local mcp as a stdio MCP server.
{
"mcpServers": {
"firecrawl": {
"command": "/path/to/firecrawl-local",
"args": ["mcp"]
}
}
}The agent gets Firecrawl’s standard MCP tools. Scrape, crawl, map, and batch scrape run on your PC, while all other tools are sent to Firecrawl Cloud.
Use —public-only When Reading Untrusted Sites
When an AI agent reads a web page, instructions hidden in the page could trick it into opening URLs on your internal network. This is known as prompt injection.
To guard against this, we recommend registering agents that handle untrusted sites with the --public-only flag.
{
"mcpServers": {
"firecrawl": {
"command": "/path/to/firecrawl-local",
"args": ["mcp", "--public-only"]
}
}
}By design, scrape / crawl / mcp can access your own PC and your internal network. With --public-only, access to localhost and private addresses is denied. Note that cloud metadata addresses and similar targets are always blocked in every mode.
Firecrawl Local vs. Self-Hosted vs. Cloud
Firecrawl already offered “self-hosted” and “Cloud” options. Firecrawl Local is the third option alongside them.
| Firecrawl Local | Self-hosted | Firecrawl Cloud | |
|---|---|---|---|
| Architecture | A single binary | Multiple services via Docker Compose (api / playwright-service / redis / rabbitmq / nuq-postgres) | Nothing to install |
| Ease of setup | Just cargo install | Clone the repo, create a .env, start Docker | Ready to use immediately |
| Search | Forwarded to Firecrawl Cloud | Runs within your own environment | Available |
| json / summary formats | Not supported (skipped with a warning) | Available with an LLM you provide | Available |
| Screenshots | Possible with Chrome | Requires Fire-engine (not in the standard setup) | Available |
| Job storage | In memory (cleared after 1 hour) | Persisted in a Postgres-backed queue | Available |
| Best for | Lightweight fetching on your machine or for AI agents | Long-running jobs, or setups that must stay within your own environment | Anyone who wants every feature without the setup |
In short, here’s how to choose:
- Want to try it out, or let an AI agent read the web → Firecrawl Local
- Need long-running crawls, LLM-based structured extraction, or search kept within your own environment → Self-hosted
- Want no installation or maintenance, and access to every feature → Firecrawl Cloud
Note that Firecrawl Local isn’t a replacement for the self-hosted version. It’s a separate build (apps/local in the repository). You can use both: Local for your machine and your agents, and self-hosted or Cloud for production.
Things to Check When Migrating from Self-Hosted
If you’re already running the self-hosted version and want to switch your SDK’s URL to Local, check the following:
- Are you calling the v0 / v1 APIs?
- Are you using the
json/summaryformats or browseractions? - Are you checking crawl IDs older than an hour? (Local jobs live in memory and are cleared after 1 hour)
- Does your health check URL point to
/v0/health/readiness? (Local uses/health)
Limitations and Things to Keep in Mind
Firecrawl Local is not a full Firecrawl server. It’s best to know what it can’t do up front.
-
A single binary, with no DB, queue, or Docker required
-
scrape / crawl / map / batch scrape are free and need no API key
-
`firecrawl-local init` registers MCP with major AI agents in one go
-
Existing Firecrawl SDKs and MCP setups work with just a URL change
-
With `FIRECRAWL_CLOUD=off`, nothing is ever sent to the cloud
-
The `json` and `summary` formats are unsupported (they require an LLM)
-
Browser `actions` are ignored
-
Crawl / batch jobs live in memory and are cleared after 1 hour, or whenever you stop `serve`
-
Search, blocked-page retries, and Alexandria require the cloud
-
v0 / v1 APIs are not supported
-
Installation requires Rust and a C compiler
The license is stated as AGPL-3.0. For personal use, this is unlikely to be an issue, but if you plan to modify it and build it into a service you offer, review the license terms beforehand.
Who Is Firecrawl Local Best For?
Good fit for:
- People who want AI agents (Claude Code, Cursor, etc.) to read documentation sites
- People who want to scrape casually on their own machine, without pay-as-you-go pricing or server maintenance
- People who don’t want to send URLs to external servers, such as when working on client projects
- People who want to run existing Firecrawl SDK code locally as-is
Not a good fit for:
- People who need long-running crawls that won’t be interrupted
- People who need LLM-based structured extraction (
json/summary) - People who want search to run entirely within their own environment (self-hosting is a better fit)
- People who’d rather not install Rust or a C compiler (Cloud is easier)
Frequently Asked Questions
Is Firecrawl Local free?
Yes. scrape / crawl / map / batch scrape run locally and don’t need an API key. Search, blocked-page retries, and Alexandria use the cloud. Alexandria is available if you log in with a free account.
Do I need Docker?
No. It runs from a single binary. However, if you want to render JavaScript pages inside a container, you’ll need a dedicated seccomp profile to run Chrome’s sandbox.
Can I use it without the cloud?
Yes. With FIRECRAWL_CLOUD=off, scrape / crawl / map still run on your own PC as usual. In exchange, search, blocked-page retries, and Alexandria become unavailable.
Can it replace the self-hosted version?
It depends on your use case. For lightweight fetching on your machine or for agents, Local is enough. But if you need long-running jobs or LLM-based extraction, you’ll still need the self-hosted version.
Which AI tools does it work with?
firecrawl-local init supports Claude Code / Claude Desktop / Cursor / Windsurf / Gemini CLI / Codex. Any other MCP-compatible tool can use it too, by registering firecrawl-local mcp as a stdio MCP server.
Summary
Firecrawl Local lets you bring Firecrawl’s scraping features straight to your own PC in a single binary.
- Get started with just
cargo install firecrawl-local - scrape / crawl / map / batch scrape are free and need no API key
- A single
initregisters it as an MCP server with major AI agents - Cloud features like search can be toggled with
FIRECRAWL_CLOUD json/summaryand long-running jobs aren’t supported, so use self-hosted or Cloud if you need them
To get started, try fetching a documentation site you’ve been curious about.
