Best web data MCP servers in 2026: Exa, Firecrawl, Apify, Perplexity

Top MCP servers for web data

The 6 best MCP servers for pulling information from the web, research papers and Hugging Face, for your own research in Claude Code and inside your product, with prompts and code

These 6 MCP servers get information from outside your company: the open web, social platforms, research papers and Hugging Face. You can use them two ways. In Claude Code, they do your own research for marketing, product and development. Inside your product, you attach the same hosted servers to a Claude API or AI SDK call and your agent gets the tools.

We picked them by GitHub stars and visitors on PulseMCP in October 2026. PulseMCP is a public directory of MCP servers that estimates how many people visit each one every week.

#	MCP	Gets you	Free tier	Then
1	Exa	Web search built for AI	$10/month credit	$4–7 per 1K searches
2	Firecrawl	Any page or site as clean markdown	1K credits/month	from $16/month
3	Apify	5,000+ ready-made scrapers	$5/month usage	from $19/month
4	Perplexity	Cited answers and deep research	None	Pay per use
5	arXiv	Research papers, section by section	Free	Free
6	Hugging Face	Models, datasets, papers and Spaces	Free	Free

1. Exa

Exa is a search engine built for AI agents. It returns full page content, not ten blue links, and understands requests like "B2B companies that launched an MCP server this quarter". It works without a key for testing.

We gave Claude Exa's hosted server, the same URL you'd attach in your backend, and asked for SaaS companies that launched an MCP server in the last 90 days. Four searches later it had 8, each with a date and the vendor's own announcement:

Terminal

Loading editor...

claude mcp add --transport http exa https://mcp.exa.ai/mcp

Build it into:

a CRM that researches every new lead: company, funding, recent news

a recruiting tool that finds candidate profiles and their work

a writing app that pulls cited sources into drafts

Hosted URL: https://mcp.exa.ai/mcp

Watch out: its agent_run tool costs up to $1 per call at high effort. Cap it, or allow only web_search_exa and web_fetch_exa with ?tools=.

Alternatives: Tavily for search, extract and crawl in one server, or Brave Search for an independent index at a low price. Full list on the Exa page.

2. Firecrawl

Firecrawl turns any URL into clean markdown or structured JSON: one page, a whole site, or search results with their content. It handles JavaScript-heavy pages and pages that block simple scrapers.

We asked it to compare the pricing pages of Linear, Asana and ClickUp. Four scrapes later Claude had the prices, the one limit each uses to force an upgrade, and seven ideas for a pricing page that does better:

Terminal

Loading editor...

claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp

Build it into:

competitive intel that checks rivals' pricing pages every week

a support bot that learns a customer's docs site on signup

an ecommerce tool that extracts product specs into JSON

Hosted URL: https://mcp.firecrawl.dev/v2/mcp

Watch out: a crawl can flood the model's context. Set a page limit and depth. JSON extraction costs 5 credits per page instead of 1.

Alternatives: Crawl4AI, a free open-source crawler you run yourself, or Jina AI for reading single pages with a generous free tier. Full list on the Firecrawl page.

3. Apify

Apify's store has 5,000+ ready-made scrapers, called Actors, for Google Maps, LinkedIn, Instagram, TikTok, Amazon, G2 and more. The MCP lets your agent find and run them, so you don't write or maintain scrapers.

Terminal

Loading editor...

claude mcp add --transport http apify https://mcp.apify.com

Build it into:

lead generation from Google Maps businesses

social analytics that pulls Instagram or TikTok profiles on demand

review monitoring across G2 and app stores

Hosted URL: https://mcp.apify.com?tools=apify/google-search-scraper

Watch out: pin the Actors you trust with ?tools=. Left to discover on its own, the agent may pick an expensive or low-quality community Actor.

Alternatives: Bright Data when a site blocks scrapers and you need proxies built in, or Oxylabs if you already use its proxies. Full list on the Apify page.

4. Perplexity

Perplexity's official server gives Claude web search, written answers with sources, reasoning and a deep research mode, all through the Perplexity API. Use it when you want a cited answer or a sourced report, not a list of pages.

It's also the quickest way to see what an AI answer engine says about your category: who it recommends, which sources it cites and whether you're in them.

Terminal

Loading editor...

claude mcp add --transport http perplexity https://api.perplexity.ai/mcp --header "Authorization: Bearer YOUR_API_KEY"

Hosted URL: https://api.perplexity.ai/mcp

Watch out: it needs a Perplexity API key with credits, separate from a Perplexity Pro subscription. Deep research calls cost more and take minutes.

5. arXiv

The arXiv server turns Claude into a research assistant for papers. It searches arXiv by topic, category and date, reads papers one section at a time, follows citations both ways through Semantic Scholar and exports BibTeX.

It's a community project that runs locally and needs no API key. Use it when a feature, a model choice or an article depends on what the research actually says.

Terminal

Loading editor...

claude mcp add --transport stdio --scope user arxiv -- uvx arxiv-mcp-server

Watch out: papers on arXiv aren't peer reviewed by default, so ask the agent to say when a result comes from a preprint. It runs locally only, so it's a research tool for your team, not one to attach to your product.

6. Hugging Face

Hugging Face's official server connects Claude to the Hub: models, datasets, Spaces, papers and the docs. Ask for "a small model that classifies support tickets" and it comes back with links, downloads and licenses.

It can also run community AI apps. Many Gradio Spaces work as MCP tools, like speech to text or background removal. The server is hosted by Hugging Face and free to connect.

Terminal

Loading editor...

claude mcp add hf-mcp-server -t http https://huggingface.co/mcp?login

Then run /mcp and sign in with your Hugging Face account.

Watch out: community Spaces can be slow, sleeping or gone, and each one you add brings more tools into context. Check a model's or dataset's license before you build on it.

Use them inside your product

Exa, Firecrawl, Apify and Perplexity are hosted, so the server you use in Claude Code also works in your backend. Pass its URL in your API call. Claude connects to it, picks the tools it needs and returns the answer.

Claude API:

Python

Loading editor...

import anthropic
client = anthropic.Anthropic()

response = client.beta.messages.create(
    model="claude-opus-5-5",
    max_tokens=2048,
    betas=["mcp-client-2025-11-20"],
    mcp_servers=[{
        "type": "url",
        "url": "https://mcp.exa.ai/mcp",
        "name": "exa",
        "authorization_token": EXA_API_KEY,
    }],
    tools=[{"type": "mcp_toolset", "mcp_server_name": "exa"}],
    messages=[{"role": "user", "content": "Research acme.com before my sales call"}],
)

Vercel AI SDK:

TypeScript

Loading editor...

import { createMCPClient } from '@ai-sdk/mcp';
import { generateText } from 'ai';

const mcp = await createMCPClient({
  transport: {
    type: 'http',
    url: 'https://mcp.firecrawl.dev/v2/mcp',
    headers: { Authorization: `Bearer ${process.env.FIRECRAWL_API_KEY}` },
  },
});

try {
  const { text } = await generateText({ model, tools: await mcp.tools(), prompt: 'Summarize acme.com/pricing' });
} finally {
  await mcp.close();
}

⚠️

The Claude API sends your key only as a Bearer token and can't add custom headers. Exa, Firecrawl, Apify and Perplexity work as is.

Before you ship:

Allow only the tools you need. Every server lets you allowlist tools. Fewer tools means lower cost and fewer surprises.

Treat web content as untrusted. Pages from Exa, Firecrawl and Apify can carry instructions aimed at your agent. Keep write tools behind human approval when both are in the same session.