Round-up · 8 options

The best MCP servers for web browsing

An agent that can't read a JavaScript-rendered page can't read most of the web. Here are the MCP servers that give it a browser, how they run, and what each costs.

By Roger Campos · Last updated: September 2026

TL;DR

The reference fetch server is free but has no browser, so single-page apps come back empty. Playwright MCP drives a real browser on your machine — best when the agent has to click. For reading pages with nothing to install, use a hosted server: URLpipe (14 tools, JavaScript rendered, read-only tokens), Firecrawl (search and crawl), Jina (read and search) or Apify (Actors).

Start free See the API

Free plan, no credit card. 1,000 credits a month.

The shortlist

8 options, and what each is best for

1. URLpipeThat's us
Docs
Per-URL web data API
A hosted MCP server at https://urlpipe.dev/mcp with 14 tools: every URLpipe endpoint, plus results, request history, projects and usage.
Best for
Agents that read, screenshot and audit JavaScript-rendered pages with nothing installed.
Watch out for
Reads pages; it doesn't click, type or log in. Tools spend credits like the HTTP API.
To start
$0 — 1,000 credits/mo, no card · Starter $19/mo — 20,000 credits
2. Fetch MCP server
github.com
Reference MCP server
The Model Context Protocol project's reference fetch server: one tool that downloads a URL over plain HTTP and converts it to Markdown.
Best for
Static pages and documentation, free, run on your own machine with uvx.
Watch out for
No browser, so JavaScript-rendered pages come back as an empty shell. Its README warns it can reach local and internal addresses.
To start
Free — open source (MIT)
3. Playwright MCP
github.com
Local browser MCP server
Microsoft's MCP server that drives a real browser on your machine, handing the model accessibility snapshots to navigate, click, type and screenshot with.
Best for
Agents that have to act on a page — log in, fill a form, click through — rather than just read it.
Watch out for
It runs a browser locally and the model steers it step by step, with a snapshot in context at each step — heavier than reading a page once as Markdown.
To start
Free — open source (Apache-2.0)
4. Firecrawl
firecrawl.dev
Crawl-and-extract API
Firecrawl's MCP server, hosted at mcp.firecrawl.dev or run with npx, with scrape, search, map, crawl and agent tools.
Best for
Agents that search the web and crawl sites, not just read one page.
Watch out for
Tools spend Firecrawl credits, cached results included; the keyless tier has only scrape, search and parse.
To start
$0 — 1,000 credits/mo, 2 concurrent browsers · Hobby $16/mo annual, $19 monthly — 5,000 credits
URLpipe vs Firecrawl
5. Jina Reader
jina.ai
URL-to-Markdown reader
Jina's hosted MCP server at mcp.jina.ai with 12 tools: read a URL, screenshot it, search the web, arXiv and more.
Best for
Agents that search as well as read, and read PDFs.
Watch out for
Most tools need a Jina key; reading is billed in tokens.
To start
20 requests/min with no key; 10M tokens for every new key
URLpipe vs Jina Reader
6. Apify
apify.com
Scraper platform
Apify's hosted MCP server at mcp.apify.com, exposing Actors — including its RAG Web Browser — as tools.
Best for
Agents that need a ready-made scraper for a specific site.
Watch out for
Actors bill by compute, so a tool call's cost varies.
To start
$0 — $5 of platform credit/mo · Starter $19/mo ($17 annual)
Apify alternatives
7. ScreenshotOne
screenshotone.com
Screenshot API
ScreenshotOne's MCP server, hosted with OAuth or run with npx, for screenshots and Markdown.
Best for
Agents that mostly need to see a page.
Watch out for
Screenshots and Markdown only; each call is a render on your plan.
To start
$0 — 100 screenshots/mo, no card · Basic $17/mo — 2,000 screenshots
URLpipe vs ScreenshotOne
8. Browserless
browserless.io
Hosted headless browsers
Browserless's hosted MCP server with a browser agent, smart scrape, Lighthouse and crawl tools.
Best for
Agents that need a remote browser to act in.
Watch out for
Billed in 30-second units of browser time.
To start
$0 — 1,000 units/mo, 2 concurrent · Prototyping $25/mo annual — 20,000 units

The problem

Why the reference fetch server isn't enough

The Model Context Protocol's reference fetch server downloads a URL over plain HTTP and converts the HTML to Markdown. That works for static pages and documentation. On a single-page app, the HTML the server sends is an empty container and a script bundle — the content arrives after JavaScript runs, and there is no browser to run it. Its README also warns that it can reach local and internal addresses.

Two shapes

Local browser or hosted API?

Local browser servers — Playwright MCP — run a real browser on your machine. The model sees accessibility snapshots and steers step by step: navigate, click, type. That's what you want when the agent has to act on a page, and it costs nothing but your CPU and the tokens each snapshot takes.

Hosted servers — URLpipe, Firecrawl, Jina, Apify, ScreenshotOne, Browserless — render the page on their side and hand back the result: Markdown, a screenshot, metadata. Nothing to install, a bearer token in the config, and the page arrives in one tool call.

Setup

Connecting URLpipe

URLpipe's server speaks streamable HTTP. Most clients take a JSON config of this shape; see the MCP docs for each client. An organization token can be limited to one project, set to expire, or made read-only — able to read results already paid for, but unable to spend credits.

MCP client config

{
  "mcpServers": {
    "urlpipe": {
      "type": "http",
      "url": "https://urlpipe.dev/mcp",
      "headers": { "Authorization": "Bearer YOUR_ORGANIZATION_TOKEN" }
    }
  }
}

Tell the agent to use fetch_markdown to read a page — it is the cheapest tool at 1 credit — and a repeat read of the same URL within max_age is free.

Recommendations

Which one to pick

If your agent needs to…	Pick
Read static docs, for free	Fetch (reference server)
Click, type and log in	Playwright MCP
Read, screenshot and audit JavaScript pages, nothing installed	URLpipe
Search the web and crawl sites	Firecrawl
Search and read, including PDFs	Jina
Run a ready-made scraper for a specific site	Apify

Compared

Every option, side by side

Capability	URLpipe	Fetch MCP server	Playwright MCP	Firecrawl	Jina Reader	Apify	ScreenshotOne	Browserless
MCP server	Hosted	Local (uvx)	Local (npx)	Hosted, or npx	Hosted	Hosted	Hosted, or npx	Hosted
Renders JavaScript (real browser)	Yes	No	Yes	Yes	Yes	Yes	Yes	Yes
Clean Markdown from a URL	Yes	Yes	No	Yes	Yes	Yes	Yes	Yes
Screenshots	Yes	No	Yes	Yes	Yes	Partial	Yes	Yes
Metadata / Open Graph	Yes	—	—	Yes	Partial	Partial	Yes	Partial
Lighthouse audits	Yes	—	—	No	No	—	No	Yes
JavaScript console capture	Yes	—	—	—	No	—	No	Your own code
Whole-site crawling	No	—	—	Yes	No	Yes	No	Beta
Self-hostable	No	MIT	Apache-2.0	AGPL-3.0 core	Apache-2.0 repo	—	No	SSPL-1.0 or commercial
Free tier	1,000 credits/mo	Free, open source	Free, open source	1,000 credits/mo	10M tokens per key	$5 credit/mo	100 renders/mo	1,000 units/mo
EU processing, AI included	Yes	—	—	—	Experimental EU endpoint	—	—	—

Fetch MCP server: checked against github.com/modelcontextprotocol/servers/tree/main/src/fetch on September 24, 2026. Playwright MCP: checked against github.com/microsoft/playwright-mcp on September 24, 2026. Firecrawl: checked against firecrawl.dev/pricing, docs.firecrawl.dev/billing, docs.firecrawl.dev/features/scrape, docs.firecrawl.dev/webhooks/overview, docs.firecrawl.dev/mcp-server, firecrawl.dev/blog/introducing-alexandria-series-b on September 24, 2026. Jina Reader: checked against jina.ai/reader, jina.ai/api-dashboard/pricing, github.com/jina-ai/reader, github.com/jina-ai/MCP, huggingface.co/jinaai/ReaderLM-v2 on September 24, 2026. Apify: checked against apify.com/pricing, apify.com/apify/website-content-crawler, apify.com/apify/rag-web-browser, docs.apify.com/platform/integrations/mcp on September 24, 2026. ScreenshotOne: checked against screenshotone.com/pricing, screenshotone.com/docs/options, screenshotone.com/integrations/mcp, screenshotone.com/privacy-policy on September 24, 2026. Browserless: checked against browserless.io/pricing, docs.browserless.io/rest-apis/performance.md, docs.browserless.io/rest-apis/smart-scrape.md, docs.browserless.io/mcp/overview.md, github.com/browserless/browserless on September 24, 2026. Products change — confirm on each vendor's site before you decide.

FAQ

Frequently asked questions

What is the best MCP server for web browsing?

For reading JavaScript-rendered pages with nothing to install, a hosted server such as URLpipe, Firecrawl or Jina. For an agent that must click and type, Playwright MCP. For static pages only, the free reference fetch server.

Does the reference fetch MCP server render JavaScript?

No. It fetches over plain HTTP and converts the HTML, so pages that build their content with JavaScript come back mostly empty.

Can an MCP server take screenshots?

Yes. URLpipe, Jina, ScreenshotOne, Browserless and Playwright MCP all have screenshot tools. URLpipe's capture_screenshot is full page by default.

Is the Puppeteer MCP server still maintained?

No. The reference Puppeteer server was moved to the archived servers repository, which became read-only in May 2025. Playwright MCP is the maintained local option.

Keep reading

Web access for AI agents The URLpipe API as an MCP server Claude Desktop MCP server Web access for Claude and Cursor Cookie consent banner All of compare

Try URLpipe on the free plan.

1,000 credits a month, no card. Confirm your email and your key is live — most people make their first request inside five minutes.

Start free Email the team

The best MCP servers for web browsing

8 options, and what each is best for

1. URLpipeThat's us

2. Fetch MCP server

3. Playwright MCP

4. Firecrawl

5. Jina Reader

6. Apify

7. ScreenshotOne

8. Browserless