Best Web Scraping MCP Servers & Skills for Claude Code in 2026
Firecrawl, Bright Data, Zyte, and Crawlora compared as Claude Code MCP servers and Agent Skills — install commands, tool counts, and which fits which job.
Claude Code's skill and MCP-server ecosystem grew fast enough in 2026 that "which web-scraping tool do I wire in" is now a real decision with real trade-offs, not a coin flip between near-identical wrappers. Some of these plugins hand your agent clean markdown from any URL. Others are built around a fixed catalog of named platforms and return structured JSON instead. One of them writes you an actual, runnable Scrapy project. This guide compares the options that show up in Claude Code's own plugin marketplace, with verified install commands and what each one is actually built to do.
MCP server vs. Agent Skill — the short version
A skill is a SKILL.md instruction file (plus optional scripts) that Claude reads and follows — no persistent connection, no protocol overhead, install with one command. An MCP server is a running process (local or hosted) that exposes typed tools over the Model Context Protocol; Claude discovers and calls them for the length of a session. Several vendors below ship both: a lightweight skill for simple one-off use, and a full MCP server when you want the tools always available. See Give your AI agent live web data with MCP for the deeper mechanics.
What to evaluate
- Generic crawl (any URL → markdown) vs. structured platform data (named sites → JSON) — these solve different problems and most agents eventually need both.
- Tool/skill count and what they actually cover — a bigger number means nothing if the platforms you need aren't in it.
- Anti-bot handling — does it fail quietly on protected sites, or is that the vendor's actual specialty?
- Output shape — markdown for an agent to read, or a schema-typed JSON record for a pipeline.
- Pricing model — credit-based vs. subscription, and whether a failed request still gets billed.
- Install path — a single `claude mcp add` / `claude plugin install` command, or manual config file editing.
The contenders
| Tool | Format | Coverage | Install | Best for |
|---|---|---|---|---|
| Firecrawl MCP | MCP server | 20 tools — scrape/crawl/search/extract any URL | claude mcp add firecrawl --env FIRECRAWL_API_KEY=... -- npx -y firecrawl-mcp | Generic crawl-to-markdown for arbitrary sites |
| Bright Data | 21 skills + MCP | 60+ MCP tools, 40+ named platforms (Amazon, LinkedIn, TikTok...) | curl -fsSL https://cli.brightdata.com/install.sh | bash then bdata login | Anti-bot-hardened extraction at platform scale |
| Zyte Web Data | 14 skills (Claude Code plugin) | Generates a full Scrapy project from a prompt | claude plugin marketplace add zyte-ai/claude-skills | A real, deployable, testable spider — not a one-off fetch |
| Crawlora | 5 skills + MCP | 893 MCP tools / 828 skill endpoints across 81 platform groups | claude mcp add --transport http crawlora https://mcp.crawlora.net/mcp --header "Authorization: Bearer $CRAWLORA_API_KEY" | Structured JSON from named platforms, pay-on-success |
Firecrawl MCP — generic crawl-to-markdown
Firecrawl's official MCP server lists 20 tools on its Claude Code Marketplace entry (6.5k GitHub stars, verified/official publisher badge) and is built around its core product: turn any URL into clean markdown, crawl a site, or run a structured extraction with a JSON schema. It doesn't work from a fixed catalog of named platforms — point it at any public page. Choose it when the job is "read this arbitrary website," not "get me this platform's structured data." Firecrawl also ships a keyless free tier (1,000 credits/month, no account) for quick one-off use.
Bright Data — anti-bot scale across 40+ named platforms
Bright Data's brightdata/skills plugin is the largest of the four: 21 skills covering search, scrape, 40+ platform-specific data feeds (Amazon, LinkedIn, Instagram, TikTok, YouTube, Zillow, and more), a full MCP orchestration skill for its 60+-tool MCP server, proxy code generation across four network tiers, and even a competitive-intelligence skill that chains scrape + search into a structured report. Its MCP server connects over a remote URL (https://mcp.brightdata.com/mcp?token=...) with optional groups= filtering (ecommerce, social, business, finance...) so an agent isn't handed all 60+ tools at once. Choose it when anti-bot reliability on hard, named targets is the actual constraint and you're comfortable with proxy-network-style pricing.
Zyte Web Data — a real Scrapy spider, not a one-off scrape
Zyte's Claude Code plugin takes a different shape entirely: instead of returning data directly, its /scrape skill runs a five-stage pipeline (define fields → analyze the site → scaffold a Scrapy project → generate extraction code → generate the spider) and hands you a runnable project with a passing test suite (uv run scrapy crawl <spider> && uv run pytest fixtures/), optionally deployable to Scrapy Cloud for scheduled runs. It's 14 skills total, packaged as an official Claude Code plugin (zyte-ai/claude-skills). Choose it when you need a maintained, inspectable, redeployable spider — not a single JSON response — and you're comfortable owning a Scrapy codebase afterward.
Crawlora — structured JSON across 81 platform groups, pay-on-success
Crawlora ships both an MCP server (893 tools, hosted at mcp.crawlora.net/mcp, one-command install via claude mcp add) and a lighter skills package (crawlora-skills, 5 skills: an umbrella catalog skill plus focused ones for product/price research, YouTube, app-store reviews, and SERP/keyword research) that installs with npx skills add or as a Claude Code plugin marketplace (/plugin marketplace add Crawlora-org/crawlora-skills). Every tool maps to one documented REST endpoint and returns normalized JSON — not markdown to re-parse — across search, maps, e-commerce, social, finance, travel, jobs, media, app stores, and reviews. Billing is pay-on-success: a failed (non-2xx) call isn't charged, and the free tier is 2,000 credits/month with no card. Choose it when the job is structured records from named platforms and you don't want to write or maintain a parser per source.
How to choose, by job
- You need to crawl or read arbitrary websites for agent context → Firecrawl MCP.
- You need reliable extraction from 40+ specific consumer/social platforms at scale → Bright Data.
- You need a real, maintainable, redeployable scraper codebase → Zyte Web Data.
- You need structured JSON from named platforms without maintaining a parser, billed only on success → Crawlora MCP or Skills.
- You're not sure yet → most production agent stacks end up running a generic crawler and a structured-platform tool side by side — they solve different halves of the same problem.
Add structured web data to Claude Code
893 tools across 81 platform groups, one API key, billed only on success. 2,000 free credits a month, no card.
Sources
Related reading
- Crawlora vs Firecrawl — structured platform APIs vs. AI-native crawl/scrape/map.
- Crawlora vs Bright Data — developer-first API catalog vs. a broad enterprise proxy/dataset platform.
- Crawlora vs Zyte — structured endpoints vs. a Scrapy-based crawling stack.
- Give your AI agent live web data with MCP — how MCP works and how to connect in three minutes.
- Best Web Search APIs for AI Agents in 2026 — the search half of the same problem.
- Best AI Web Scraping Tools in 2026 — the broader tool landscape beyond Claude Code specifically.
Frequently asked questions
What's the best MCP server for web scraping with Claude Code?
It depends on the job. Firecrawl MCP is the pick for crawling or reading arbitrary websites. Bright Data is built for anti-bot-hardened extraction across 40+ named platforms at scale. Zyte's Web Data plugin generates an actual, deployable Scrapy spider instead of returning data directly. Crawlora fits when you want structured JSON from named platforms (search, maps, commerce, social, finance) without maintaining a parser, billed only on success.
What's the difference between a Claude Code skill and an MCP server?
A skill is a SKILL.md instruction file (plus optional scripts) Claude reads and follows on demand — no persistent connection. An MCP server is a running process that exposes typed tools over the Model Context Protocol for the length of a session. Several vendors, including Crawlora, Bright Data, and Zyte, ship both: a lightweight skill for quick one-off use and a full MCP server for always-on tool access.
How many tools does each MCP server expose?
As listed on each vendor's own repo or Claude Code Marketplace entry in August 2026: Firecrawl MCP lists 20 tools, Bright Data's MCP server exposes 60+ tools (filterable by group), and Crawlora's hosted MCP server exposes 893 tools across 81 platform groups. Zyte doesn't expose a fixed tool count the same way — its plugin runs a 5-stage pipeline that generates a Scrapy project rather than returning tool-call results directly.
Does Firecrawl or Crawlora have a better free tier for testing?
Both offer credit-based free tiers with no card required: Firecrawl Keyless gives 1,000 credits/month, and Crawlora gives 2,000 credits/month. Firecrawl's free tier works without creating an account at all; Crawlora's requires a free account but bills only successful (2xx) responses even on paid plans.
Can I use more than one of these at once?
Yes, and many production agent setups do — a generic crawler (Firecrawl or Bright Data) for arbitrary sites, plus a structured-platform tool (Crawlora) for the handful of named sources an agent hits constantly, so you're not maintaining a custom parser for platforms that already have a documented endpoint.
How do I install Crawlora's MCP server in Claude Code?
Run claude mcp add --transport http crawlora https://mcp.crawlora.net/mcp --header "Authorization: Bearer $CRAWLORA_API_KEY" with a free API key from crawlora.net (2,000 credits/month, no card). A lighter alternative is the skills package: npx skills add github.com/Crawlora-org/crawlora-skills --all, or add it as a Claude Code plugin marketplace with /plugin marketplace add Crawlora-org/crawlora-skills.