Firecrawl Review 2026: The Web Scraping API I Actually Use

I have been hammering on Firecrawl for the better part of a year now — first as a curious side experiment, then as the default scraping layer for almost every retrieval-augmented generation (RAG) project that crosses my desk. This is my honest, hands-on review for 2026.

If you have ever tried to feed an LLM raw HTML, you already know the pain. Half the tokens are inline styles, the other half are nav junk, and the actual content shows up somewhere around character 14,000. Firecrawl exists to delete that problem. You hand it a URL, you get back clean markdown — or a JSON object that matches a schema you defined — and the model can reason over it like a normal document.

Built by Mendable and released to the public in 2024, Firecrawl has grown faster than almost any other tool in the AI infra space. As of April 2026 the open-source repo has crossed 40,000 GitHub stars (the vendor proudly cites 97K across all forks and mirrors), it is Y Combinator backed, MIT licensed, and shipping new features almost every week. The hosted API is what most teams pay for, but you can self-host the entire stack if you want.

This review covers what Firecrawl actually does, how it compares to ScrapingBee, Apify and Bright Data, what it really costs once you turn on extraction, and whether it deserves a spot in your stack. Spoiler: for AI engineers, it almost certainly does.

Key Features

Firecrawl exposes six primary endpoints. Every other "feature" on the marketing page is a flag or an option layered on top of these.

1. Crawl — Walk an Entire Website

The flagship capability. Point Firecrawl at a root URL, set a depth limit and an include/exclude pattern, and it recursively discovers every page on the domain. The crawl handles sitemaps, robots.txt, pagination, infinite scroll and JavaScript-rendered links. You get back a job ID, poll it, and pull every page as clean markdown.

I used this to ingest the entire Convex documentation (~600 pages) in about four minutes flat. No proxy babysitting, no Selenium, no curl-and-cheerio Frankenstein.

2. Scrape — Single Page, Clean Markdown

The bread and butter. One URL in, one markdown document out. The default response strips nav, footer, ads and script tags, returns a markdown field plus optional html, links, metadata and screenshot fields. P95 latency in my testing hovered around 3.4 seconds, which matches Firecrawl's published numbers.

3. Extract — LLM-Powered Structured Data

This is where Firecrawl pulls ahead of every classic scraping API. You hand it a URL (or a list of URLs, or a wildcard like https://news.ycombinator.com/*) plus a Zod or JSON Schema, and Firecrawl runs an internal LLM to populate the fields. No selectors, no XPath, no maintenance when the page changes.

I used Extract to scrape pricing pages across 80 SaaS competitors for a positioning deck — schema in, perfect JSON out. Saved me about a week.

4. Actions — Click, Type, Wait, Screenshot

For pages that hide content behind interaction, Firecrawl exposes a small DSL of browser actions: wait, click, write, press, screenshot, scroll, scrape. You chain them in the API call and Firecrawl drives a real browser session. This is the hatch you climb through when a page needs you to dismiss a modal, log in or expand a "show more" button before the content shows up.

5. Map — Discover the Structure of a Site

Map is the cheapest endpoint. It returns every URL Firecrawl can discover on a domain — no scraping, no rendering, just the link graph. It is the right call when you want to inspect a site before you commit a full crawl, or when you only need to know which URLs exist (sitemap reconstruction, dead link checking, etc.).

6. Search — Web Search With Content

Firecrawl Search takes a query, runs a web search, then optionally scrapes each result into clean markdown in the same call. It is essentially "search + scrape in one round trip" and it competes head-on with the Exa, Tavily and Brave Search MCP servers that I have reviewed elsewhere on Skiln.

Bonus: Official MCP Server

Firecrawl ships an official Model Context Protocol server. Drop it into Claude Desktop, Cursor or Claude Code and your agent can call firecrawl_scrape, firecrawl_crawl, firecrawl_extract and firecrawl_search as native tools. This is the integration that turned Firecrawl from "useful API" into "always installed MCP" in my own setup.

How to Use Firecrawl

Setup is genuinely two minutes. Grab an API key from firecrawl.dev, install the SDK, and you are scraping.

Install

# Node / TypeScript
npm install @mendable/firecrawl-js

# Python
pip install firecrawl-py

Scrape a Single Page

import FirecrawlApp from "@mendable/firecrawl-js";

const app = new FirecrawlApp({ apiKey: process.env.FIRECRAWL_API_KEY });

const result = await app.scrapeUrl("https://skiln.co/mcps", {
  formats: ["markdown", "links"],
  onlyMainContent: true,
});

console.log(result.markdown);

The onlyMainContent: true flag is the one I leave on by default. It strips nav, footer and sidebars and saves you a remarkable amount of token budget downstream.

Crawl an Entire Site

const job = await app.crawlUrl("https://docs.firecrawl.dev", {
  limit: 200,
  scrapeOptions: { formats: ["markdown"] },
});

// job.data is an array of pages, each with markdown + metadata

Extract Structured Data

import { z } from "zod";

const schema = z.object({
  productName: z.string(),
  price: z.number(),
  currency: z.string(),
  features: z.array(z.string()),
});

const result = await app.extract(
  ["https://example.com/pricing"],
  {
    prompt: "Extract the headline product, its price and key features.",
    schema,
  }
);

Hook It Into Claude as an MCP

claude mcp add firecrawl -- npx -y firecrawl-mcp

Then export FIRECRAWL_API_KEY in your environment and Claude will see firecrawl_scrape, firecrawl_crawl, firecrawl_extract and firecrawl_search as native tools. You can find this and other useful servers on the Skiln MCP directory.

Pricing

Firecrawl's pricing model is credit-based. One credit equals one scraped page at the base level — but turn on JSON extraction or "Enhanced Mode" and a single page can cost up to nine or ten credits. Read the fine print before you commit.

A note on math: I ran a 100-page Standard-plan crawl with Extract enabled and burned roughly 850 credits, not 100. If you are budgeting for an extraction-heavy workload, divide your credit count by ten and treat that as your real page budget. With a plain markdown crawl (no extraction) the 100,000-credit Standard plan gets you very close to the full 100,000 pages.

Credits do not roll over month-to-month except for auto-recharge add-ons and custom annual contracts. If you have spiky workloads, set a hard budget cap in the dashboard.

Pros and Cons

Best Alternatives

In practice, I run Firecrawl as the default and reach for Bright Data only when a target is hostile enough that Firecrawl's built-in proxies cannot crack it. Apify is the right call when there is already a community actor for the exact site I need (Amazon, LinkedIn, Google Maps). For pure search-and-summarise I lean on Tavily — and you can read my Exa vs Tavily vs Brave Search MCP comparison for the deep dive on those.

FAQ

What is Firecrawl?

Firecrawl is an open-source web scraping and crawling API built specifically for LLMs. It turns any website into clean markdown or structured JSON ready for RAG, agents and fine-tuning datasets.

Is Firecrawl free?

Yes. The hosted API ships with a 500-credit free tier, no credit card required. The MIT-licensed open-source repo is free to self-host on your own infrastructure.

How much does Firecrawl cost in 2026?

Hobby is $16/month for 3,000 credits, Standard is $83/month for 100,000 credits, Growth is $333/month for 500,000 credits. Scale and Enterprise are custom-quoted.

Does Firecrawl handle JavaScript-rendered pages?

Yes. It runs a real headless browser, rotates proxies, and renders SPAs and infinite scroll out of the box. You only need a paid proxy upgrade for the most aggressive anti-bot targets.

What is the difference between Firecrawl crawl and scrape?

Scrape pulls a single page. Crawl recursively walks an entire site within your include/exclude rules. Map returns just the URL graph. Extract pulls structured fields against a schema you define.

Does Firecrawl have an MCP server?

Yes, an official one. Install with claude mcp add firecrawl -- npx -y firecrawl-mcp, set FIRECRAWL_API_KEY, and Claude or Cursor will see scrape, crawl, extract and search as native tools.

What are the best Firecrawl alternatives?

ScrapingBee for SERP and simple JS pages, Apify for marketplace actors, Bright Data for enterprise proxy infrastructure, and Tavily for AI search. Firecrawl wins anywhere the output is going straight into an LLM.

Final Verdict

After roughly a year of running Firecrawl in production across RAG pipelines, agent loops and one-off scraping jobs, my verdict is clear: this is the default web scraping API for AI engineers in 2026. The combination of clean markdown output, schema-based Extract, an official MCP server and a $16 entry price is hard to beat. I lose maybe one job in fifty to bot detection, and that is what Bright Data is for.

Is it perfect? No. Extract mode chews credits faster than the marketing page suggests, and the lack of credit rollover stings if your scraping is bursty. But the engineering quality is high, the docs are unusually good for a tool that moves this fast, and the team is shipping the right features in the right order.

If you are building anything that needs fresh content from the open web — RAG, an agent, a pricing monitor, a research bot, a competitor watcher — start with Firecrawl. You will save a week of plumbing on day one.

Rating: 4.5 / 5