How to Install the Firecrawl Skill in Claude Code (2026)

Plain WebFetch in Claude Code grabs one page and hands you the raw HTML, scripts and nav menus included. Point it at a JavaScript-heavy docs site and you often get an empty shell. The Firecrawl skill changes that math: it lets Claude crawl an entire site, render the JavaScript, and return clean markdown that's ready to reason over. Ask it to read all of a product's API docs, and it will walk every page instead of stopping at the first one.

This guide covers what Firecrawl does, how it compares to the tools you already have, and how to wire it into Claude Code step by step. It also covers the part most tutorials skip: how to handle your API key and how to check a third-party skill before you trust it.

What Firecrawl Actually Does

Firecrawl is a web-scraping and crawling API. You give it a URL and it returns content an LLM can read without choking on markup. It exposes a handful of operations, and each one maps to a different job:

• Scrape — pull a single page and return it as clean markdown or structured JSON.
• Crawl — start at one URL and follow links across a whole site, returning every page.
• Map — list all the URLs on a site quickly, without fetching the full content of each.
• Extract — pull specific structured fields (prices, author names, dates) using a schema you describe in plain language.
• Search, run a web search and return scraped results in one step.

The payoff is that Firecrawl handles the messy parts itself. It renders pages that need a browser, strips boilerplate, follows pagination, and respects rate limits. The skill, or the underlying MCP server, wires those operations into Claude Code so the agent can call them mid-task. You ask a question, and Claude decides whether to scrape one page or crawl two hundred.

Firecrawl is open-source, which matters for two reasons. You can read exactly what it does at github.com/mendableai/firecrawl, and you can self-host it if you'd rather not send URLs through a hosted service. The official docs live at firecrawl.dev.

Firecrawl vs WebFetch vs Playwright MCP

Before you install anything, it helps to know when each tool earns its place. Claude Code ships with WebFetch already. Playwright MCP is a separate browser-automation server. Firecrawl sits between them.

The short version: WebFetch is fine for a quick lookup. Playwright is what you want when a task needs clicking through a login or filling a form. Firecrawl is the tool for reading a lot of content cleanly. If you're weighing browser automation, our Playwright MCP install guide walks through that path.

Step 1: Get a Firecrawl API Key

Go to firecrawl.dev and create an account. Once you're in, open the dashboard and find the API keys section. Copy your key. It usually starts with a recognizable prefix so you can spot it later.

The free tier gives you a monthly credit allowance, which is enough to test the setup and run small crawls. If you expect heavy use, check the pricing page before you start, because a deep crawl of a large site can burn through credits fast. Self-hosting from the GitHub repo removes the per-request cost entirely, at the price of running your own instance.

Don't paste the key anywhere public yet. The next step puts it somewhere safe.

Step 2: Store Your API Key as an Environment Variable

Never hardcode the key into a config file you might commit to git. Set it as an environment variable instead.

On macOS or Linux, add this to your shell profile:

export FIRECRAWL_API_KEY="fc-your-key-here"

On Windows PowerShell:

$env:FIRECRAWL_API_KEY = "fc-your-key-here"

To make it permanent on Windows, set it through the System Properties environment variables dialog or with setx. Restart your terminal so the variable loads. This keeps the secret out of your project files and out of your git history, which is the single most common way people leak API keys.

Step 3: Add the Firecrawl MCP Server to Claude Code

The skill leans on Firecrawl's MCP server, so install that first. Claude Code has a built-in command for adding MCP servers. The pattern looks like this:

claude mcp add firecrawl --env FIRECRAWL_API_KEY=$FIRECRAWL_API_KEY -- npx -y firecrawl-mcp

That command registers a server named firecrawl, passes your API key through the environment, and runs the published MCP package with npx. The exact package name and flags can change between releases, so confirm them against the official GitHub repo and firecrawl.dev docs before you run it.

If you prefer editing config by hand, you can add the server to your Claude Code MCP settings file. The shape of the entry is roughly:

{
  "mcpServers": {
    "firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": { "FIRECRAWL_API_KEY": "fc-your-key-here" }
    }
  }
}

Use the environment-variable form where you can. Once the server is registered, restart Claude Code and run /mcp to confirm Firecrawl shows up with its tools listed.

Step 4: Add the SKILL.md (Optional but Useful)

At this point Claude can already call Firecrawl's tools. A skill adds a layer of instructions on top, telling Claude when to crawl versus scrape, how to format results, and what limits to respect. Skills live in a folder under your home directory:

~/.claude/skills/firecrawl/SKILL.md

Create that folder and drop in a SKILL.md file. A minimal one names the skill, describes when to use it, and gives Claude a few rules:

---
name: firecrawl
description: Crawl sites and extract clean markdown using the Firecrawl MCP tools.
---

When the user asks to read documentation, crawl a site, or extract
structured data from the web, use the Firecrawl tools. Prefer scrape
for one page, crawl for a whole site, and extract when the user wants
specific fields. Always respect robots.txt and the site's terms.

If you're grabbing a community skill instead of writing your own, read the whole file first. That's not optional, and the next section explains why.

A Security Note You Shouldn't Skip

A Claude Code skill is plain text, but it carries real power: it can tell Claude to run shell commands as part of a task. A skill that looks like a harmless web-scraping helper could also instruct the agent to read files or run a script you didn't expect. So before you install any third-party skill, open the SKILL.md and read every line. If it asks Claude to do something unrelated to its stated job, don't install it.

Three habits keep you safe:

• Keep the key in an environment variable, never in the skill file or a committed config. If it's already in your shell, the skill never needs to see the raw value.
• Audit before you install. Treat a SKILL.md the way you'd treat a shell script someone emailed you.
• Respect robots and terms of service. Firecrawl can crawl fast and wide. Just because the agent can pull a whole site doesn't mean a given site allows it. Check the target's robots.txt and ToS, especially for anything you plan to republish.

We wrote a fuller breakdown in our Firecrawl skill review if you want the deeper read on trust and limits.

Usage: Prompts to Try Once It's Installed

With the server live, you talk to Claude in plain language. It decides which Firecrawl operation fits. A few prompts that map cleanly to real work:

Read and summarize a docs section:

> Crawl docs.stripe.com and summarize how the authentication flow works, with code examples for the API key header.

Claude will crawl the relevant pages, pull the clean markdown, and write the summary from the rendered content instead of a half-loaded HTML shell.

Extract structured data across pages:

> Map example-store.com, then extract the product name, price, and stock status from every product page into a table.

This uses map to list URLs, then extract with a small schema. You get a table, not a wall of HTML.

Research with search:

> Search for recent posts about MCP server security best practices and give me the three most useful, with the source link for each.

Firecrawl's search runs the query and scrapes the results in one pass, so Claude has clean text to judge rather than search-engine snippets.

When Firecrawl Is Worth It, and When It Isn't

Here's my honest take after wiring this into a few projects. Firecrawl earns its keep the moment you need more than one page or the site fights back with JavaScript. Reading an entire API reference, building a dataset from a catalog, keeping a knowledge base fresh from a docs site that updates often: that's where it shines, and where the time it saves dwarfs the credit cost.

It's overkill when you need a single static page. WebFetch already does that for free, and adding an API key plus an MCP server just to read one blog post is friction you don't need. It's also the wrong tool when the task needs interaction, like logging in or clicking through a multi-step form. That's Playwright's job, and the two tools coexist happily in the same Claude Code setup.

The rule I use: if the answer is "read a lot, cleanly," install Firecrawl. If it's "read one thing" or "click some buttons," reach for WebFetch or Playwright. If you're a data person building pipelines, our roundup of AI agent skills for data scientists puts Firecrawl in context with the rest of the toolkit.

FAQ

Do I need a Firecrawl API key to use the skill?

Yes for the hosted service. Sign up at firecrawl.dev and put the key in an environment variable. Because Firecrawl is open-source, you can also self-host from the GitHub repo and point the skill at your own instance with no per-request cost.

What's the difference between the Firecrawl skill and the Firecrawl MCP server?

The MCP server exposes Firecrawl's scrape, crawl, map, extract, and search operations as tools Claude can call. The skill is a SKILL.md file that tells Claude when and how to use those tools. Most setups install the MCP server; the skill is an optional wrapper that adds guidance.

Is Firecrawl free?

There's a free tier with a monthly credit allowance, then paid plans for higher volume. Self-hosting is free apart from your own server costs. Check firecrawl.dev for current limits, since they change.

Is it safe to install a third-party Claude Code skill?

Read the SKILL.md first. A skill can instruct Claude to run shell commands, so install only from sources you trust, keep your API key in an environment variable, and never paste secrets into the skill file.

Can Firecrawl crawl any website?

Technically it can reach most public pages, but you should respect each site's robots.txt and terms of service. Crawling and republishing content you don't have rights to can land you in legal trouble, and aggressive crawling can get your IP blocked. Stay polite and check the rules first.

The Short Version

Grab a key from firecrawl.dev, store it as an environment variable, register the MCP server with claude mcp add, and optionally drop a SKILL.md in ~/.claude/skills/firecrawl/. Read any third-party skill before you trust it. Then ask Claude to crawl a docs site and watch it return clean markdown instead of broken HTML. For one-off reads, WebFetch is still the lighter choice.