How to Connect Firecrawl to Claude Code and Cut Web Scraping Token Costs by 80%

Raw HTML Is Eating Your Token Budget — Here’s the Fix

If you’ve ever asked Claude to scrape a website and watched it burn through tokens processing cookie banners, navigation menus, and ad markup before getting to the actual content, you already know the problem. The web is not clean. And feeding raw HTML to an LLM is one of the most expensive things you can do per unit of useful information.

Firecrawl solves this. It’s an MCP connector for Claude Code that takes any URL and returns clean, structured data — no HTML soup, no ad content, no banners. The token savings are real: up to 80% cheaper compared to feeding raw HTML into Claude. You get 500 free credits when you create an account, which is enough to run meaningful experiments before spending anything. And the setup takes about five minutes.

This post walks through the full setup — from adding the custom connector in Claude Code to running a live lead generation demo that produces 20 pool cleaning businesses in Austin with name, email, website, and two facts per business, delivered as an interactive HTML document.

---

Why Raw HTML Is Such a Token Problem

When Claude reads a webpage directly, it sees everything. The <nav> with 40 links. The cookie consent modal. The footer with legal text. The sidebar ads. The tracking pixel scripts.

None of that is the content you wanted. But Claude still has to process all of it to find the parts that matter. On a modern commercial website, the signal-to-noise ratio in raw HTML can be terrible — sometimes the actual article text is 10% of what gets loaded into the context window.

This is why token management in Claude Code matters so much. Every token you spend on garbage HTML is a token you can’t spend on reasoning, output, or the next task in your session. Firecrawl’s job is to strip all of that before it ever reaches Claude.

---

Step 1: Create Your Firecrawl Account and Get Your API Key

Go to firecrawl.dev and create a free account. After signup, you’ll land in a dashboard where you can find your API key.

The free tier gives you 500 credits. Each credit corresponds roughly to one page crawl. That’s enough to run the lead gen demo below several times and still have credits left over for experimentation.

Copy your API key. You’ll need it in the next step.

---

Step 2: Add Firecrawl as a Custom Connector in Claude Code

Open Claude Code. This works in the desktop app, in VS Code with the Claude Code extension, or in the terminal — wherever you normally run it.

In the Claude Code interface, look for Manage Connectors. The exact location depends on your version, but it’s typically accessible from the settings or sidebar menu. Click it.

You’ll see a list of available connectors. Firecrawl isn’t in the default list, so you need to add it as a custom connector. Click “Add custom connector” (or similar — the label varies slightly by version).

Now you need the Firecrawl MCP server URL. Go to Firecrawl’s documentation and find the MCP server configuration section. They provide the server URL and the format for passing your API key.

The configuration will look something like this:

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

Paste your API key where indicated. Save the connector configuration.

Now you have Firecrawl available as a tool Claude can call during any session. You don’t need to do anything else to activate it — Claude will use it when you ask it to.

---

Step 3: Verify the Connection Works

Before running anything complex, do a quick sanity check. Open a new Claude Code session and ask something simple:

“Use Firecrawl to fetch the content from [any URL you want to test] and summarize what’s on the page.”

Claude will call Firecrawl, which will return clean markdown instead of raw HTML. You’ll see the difference immediately — the response will be about the actual content, not about parsing around navigation elements.

If you get an error, the most common causes are:

• API key not saved correctly (check for extra spaces or missing characters)
• Network access settings blocking the Firecrawl domain (if you’re running remote routines, make sure network access is set to “full” rather than “trusted”)
• The MCP server URL being slightly off — double-check against the current Firecrawl docs

Once you get a clean summary back, the connection is working.

---

Step 4: Run the Lead Generation Demo

This is where Firecrawl becomes genuinely useful for real work. Here’s the prompt from the source video that demonstrates what it can do:

“Use Firecrawl to find me 20 leads in the pool cleaning business in Austin. I want their name, email, website, and two interesting facts from their site. Give me the results as an interactive HTML doc.”

Paste that into Claude Code (or adapt it to whatever industry and city you’re researching).

What happens next: Claude fires off requests through Firecrawl, which pulls data from business websites and directories. Firecrawl returns clean structured content. Claude extracts the specific fields you asked for — name, email, website, two facts — and formats everything into an interactive HTML document.

A few minutes later, you have a formatted document with 20 real businesses, their contact details, and insights pulled from their actual websites. The output is clean enough to use directly or export to JSON, CSV, or Excel.

The same approach works for:

• Competitor research: Pull product pages, pricing, and feature lists from competitor sites
• Content research: Gather information from multiple sources on a topic without manually visiting each one
• Market intelligence: Scrape industry directories or review sites for patterns

The key insight is that Firecrawl handles the messy web-fetching layer so Claude can focus on the extraction and formatting work it’s actually good at.

---

What’s Actually Happening Under the Hood

When Claude Code calls Firecrawl via MCP, Firecrawl does several things before returning data:

1. Fetches the page, including handling JavaScript-rendered content (which basic HTTP requests miss)
2. Strips navigation, ads, footers, cookie banners, and other non-content elements
3. Converts the remaining content to clean markdown
4. Returns that markdown to Claude

The 80% token reduction claim comes from comparing the token count of raw HTML versus the cleaned markdown for the same page. On content-heavy pages, the ratio is often even better. On pages that are mostly navigation and boilerplate, you might see less dramatic savings — but you’ll still see savings.

This is the same principle behind using compact commands to manage context rot — the less noise in the context window, the more useful work Claude can do per session.

---

Combining Firecrawl With Other Claude Code Patterns

Firecrawl works well on its own, but it gets more interesting when you combine it with other Claude Code capabilities.

With skills: If you run lead generation or competitor research regularly, wrap the Firecrawl prompt into a skill file. Store it in .claude/skills/lead-gen/skill.md with your standard fields and output format. Then you can trigger it with a slash command instead of rewriting the prompt each time.

With routines: If you want weekly competitor monitoring, you can set up a remote routine that runs Firecrawl on a schedule. Remote routines run on Anthropic’s infrastructure with your laptop off — on the Max $200/month plan you get 15 remote runs per day, 5 on Pro. Just remember that remote routines can’t read your .env file, so your Firecrawl API key needs to go in the cloud environment variables, not a local file.

With the GWS CLI: After Firecrawl pulls and formats your lead data, you can have Claude write it directly to a Google Sheet using the Google Workspace CLI — one tool that covers Gmail, Drive, Docs, Sheets, Slides, and Calendar. The combination of Firecrawl for data gathering and GWS CLI for output storage creates a complete research-to-storage pipeline without leaving Claude Code.

If you’re building more complex multi-step workflows around this kind of data pipeline, platforms like MindStudio handle the orchestration layer: 200+ models, 1,000+ integrations, and a visual builder for chaining agents and tools together without writing the glue code yourself.

---

Token Cost Reality Check

The 80% reduction is a ceiling, not a guarantee. Here’s how to think about when you’ll see the biggest savings:

High savings: News articles, blog posts, product pages, business directories. These pages have a lot of structural HTML around relatively small amounts of actual content.

Moderate savings: Documentation sites, Wikipedia-style pages. These are already fairly content-dense, so the ratio is better to start with.

Lower savings: Pages that are mostly tables or structured data. Firecrawl still helps, but there’s less noise to strip.

For the lead generation use case specifically, you’re typically hitting business websites and local directories — exactly the kind of pages where Firecrawl’s cleanup makes the biggest difference.

If you’re already thinking carefully about how to reduce token costs in Claude Code, adding Firecrawl to any workflow that involves web data is one of the highest-leverage changes you can make.

---

A Note on Browser Automation vs. Firecrawl

You might be wondering how Firecrawl compares to browser automation with Playwright. They solve adjacent problems.

Playwright is the right tool when you need to interact with a page — fill out a form, click a button, log in, navigate through a multi-step flow. It controls a real browser.

Firecrawl is the right tool when you just need to read content from a URL. It’s faster, cheaper per operation, and returns cleaner data for extraction tasks. You don’t need to spin up a browser to scrape a business directory.

Use Playwright when the page requires interaction. Use Firecrawl when you just need the content.

---

Scaling Beyond 500 Credits

The free tier is genuinely useful for getting started and for low-volume use cases. If you’re running regular research workflows — weekly competitor monitoring, ongoing lead generation, content aggregation — you’ll want to look at Firecrawl’s paid tiers.

Before upgrading, it’s worth auditing how you’re using credits. A few things that help:

• Be specific in your prompts about which pages to crawl. Crawling a whole domain costs more credits than targeting specific URLs.
• Cache results when you can. If you’re pulling the same competitor’s pricing page weekly, store the output and only re-crawl when you need fresh data.
• Combine Firecrawl with the Karpathy-style LLM wiki approach: crawl once, store the clean markdown in your knowledge base, query it repeatedly without re-crawling.

Speaking of building on top of structured data — if you’re at the point where your Firecrawl outputs are feeding into a full application (a lead management tool, a competitive intelligence dashboard, something with a real data model), Remy takes a different approach to that build step: you write a spec in annotated markdown and it compiles into a complete TypeScript stack with backend, database, auth, and deployment. The spec stays as the source of truth; the generated code is derived output.

---

The Setup in Summary

Here’s the full path from zero to working Firecrawl integration:

1. Create a free Firecrawl account at firecrawl.dev — you get 500 credits
2. Copy your API key from the dashboard
3. In Claude Code, go to Manage Connectors → Add custom connector
4. Paste in the Firecrawl MCP server configuration with your API key
5. Test with a simple fetch request to verify the connection
6. Run the lead gen prompt (or adapt it to your use case)

The whole setup takes about five minutes. The token savings start immediately on the first crawl.

One thing worth saying directly: if you’ve been avoiding web research tasks in Claude Code because they felt expensive or unreliable, Firecrawl changes that calculation. Clean data in means better output and lower costs. It’s one of those integrations that pays for itself quickly.