How to Give AI Agents Website Screenshots with MCP (2026)

Originally published on the Site-Shot blog.

If you're building an AI agent that needs to see a web page — to drive a browser, run visual QA, or
feed a screenshot into a vision model — you need a screenshot API that the agent can call on its own.
In 2026 the clean way to wire that up is the Model Context Protocol (MCP): you point your agent at an
MCP server and it gets a screenshot tool it can invoke directly. This guide explains what MCP is, why
agents need screenshots, shows a copy-paste Claude Desktop config for Site-Shot's official MCP server,
and lists which screenshot APIs actually ship an official MCP server versus those that don't.

Disclosure: this comparison is published by Site-Shot. We have kept the numbers accurate and the
trade-offs honest — including where competitors beat us. Prices checked June 2026; confirm on each
vendor pricing page before buying.

What is MCP, and why would an agent need screenshots?

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to external tools and
data. An MCP server exposes one or more tools (each with a name, a description, and a JSON input
schema); an MCP client — Claude Desktop, Cursor, Claude Code, and a growing list of agent frameworks —
discovers those tools and lets the model call them. The agent never sees your API key or HTTP plumbing;
it just sees a tool like "capture a screenshot of this URL" and decides when to use it.

Agents reach for screenshots in three recurring patterns:

• Vision pipelines. A multimodal model can't reliably reason about a page from raw HTML. Hand it a rendered PNG and it can read layout, find a button, check a chart, or verify that a deploy looks right.
• Web automation. An agent driving a flow (login, checkout, form fill) needs to look between steps to confirm what happened before it clicks again.
• Visual QA and monitoring. Capture a page on a schedule or after a change, then diff or describe it to catch broken layouts, missing images, or unexpected content.

Why ad and cookie-banner removal cuts vision-model cost

There's a money angle specific to AI agents. Vision models bill by image tokens, and token count scales
with the image. A full-page capture cluttered with ad iframes, sticky cookie-consent overlays, and
newsletter modals is more pixels and more visual noise for the model to encode — you pay for tokens
spent encoding junk, and the noise can degrade the answer. Removing ads and cookie banners at capture
time (Site-Shot's no_ads and no_cookie_popup parameters, and equivalents on most competitors) gives
the model a cleaner, often smaller image: lower token cost per call and fewer wrong answers from a popup
covering the content. For an agent that captures thousands of pages, that adds up.

Setting up Site-Shot's MCP server (copy-paste config)

Site-Shot ships an official MCP server published as site-shot-mcp
(registry id io.github.site-shot/site-shot-mcp). It runs over npx, so there's nothing to install
globally. Drop this into your Claude Desktop config file (claude_desktop_config.json —
Settings → Developer → Edit Config), restart Claude Desktop, and the tools appear:

{
  "mcpServers": {
    "site-shot": {
      "command": "npx",
      "args": ["-y", "site-shot-mcp"],
      "env": {
        "SITESHOT_API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

The server exposes two tools:

• capture_screenshot — a viewport screenshot of a URL, with the usual options (dimensions, device emulation, format, ad/cookie removal, country proxy).
• capture_full_page — a full-length capture that scrolls and stitches the entire page (up to 20,000px tall on Site-Shot), ideal for handing a whole article or dashboard to a vision model.

The same config shape works in other MCP clients (Cursor, Claude Code, etc.) — they all read a
command / args / env block. Under the hood each tool is a single GET request to Site-Shot's real
Chromium renderer, returning a PNG/JPEG or JSON (base64 image plus the target's response headers).

Honest scoping: the MCP server requires an API key. Site-Shot's free, no-signup offering is the
browser-based online tool — great for one-off captures, but it does not
issue an API key. To use the MCP server (or the HTTP API at all) you need an account and a key from the
pricing page, starting at $5/mo for 2,000 screenshots. See the
AI agents page for the full agent setup.

Which screenshot APIs ship an official MCP server?

Plenty of screenshot APIs exist; far fewer maintain an official, first-party MCP server you can trust
to track their feature set. This table sorts by that distinction.

API	Official MCP server?	Entry plan	~Cost / 1,000	Ad + cookie removal
Site-Shot	✅ Official (site-shot-mcp)	$5/mo · 2,000	$2.50 → $0.80 at scale	✅
ScreenshotOne	✅ Official	$17/mo · 2,000	$8.50	✅
Urlbox	✅ Official	$19/mo · 2,000	$9.50	✅
Microlink	✅ Official	$49/mo · 46,000	$1.07	✅
CaptureKit	✅ Official	$7/mo · 1,000	$7.00	✅
ApiFlash	⚠️ 3rd-party wrappers only	$7/mo · 1,000	$7.00	✅
ScreenshotAPI.net	❌ None	$9/mo · 1,000	$9.00	✅
Thum.io	❌ None	$2 / 10,000 hits	$0.20	❌

Cost per 1,000 is from each provider's cheapest paid plan. For an agent, "official MCP" matters: a
first-party server stays in sync with the API and is maintained by the vendor; a third-party wrapper can
lag or vanish.

The MCP-ready options, one by one

Site-Shot

Site-Shot's official server (npx -y site-shot-mcp) exposes
capture_screenshot and capture_full_page, backed by real Chromium with device emulation, country
proxies (IP, language, time zone, geolocation on every plan), full-page to 20,000px, and ad/cookie
removal that keeps vision-model images clean. It's the lowest entry commitment with an official MCP —
$5/mo for 2,000 shots, dropping to $0.80/1,000 at scale — and the only one here with a genuinely free,
no-signup browser tool to prototype against. Best for: agents that want full-featured real-browser
captures without premium per-shot pricing. Gap: no native client SDKs yet — the HTTP API is a plain
GET, so for non-MCP code you call it directly.

ScreenshotOne

ScreenshotOne ships an official MCP server (the render-website-screenshot
tool, env SCREENSHOTONE_API_KEY) alongside the deepest feature set in this list: SDKs in seven
languages, PDF and video rendering, JS/CSS injection, S3 upload, webhooks, and a dedicated AI-agents
page. It is genuinely the most polished, integration-rich option — and the priciest per shot at
$8.50/1,000 ($17/mo for 2,000; geo proxy is gated to its Growth and Scale tiers). Best for: teams that
want the richest agent tooling and don't mind paying for it.

Urlbox

Urlbox maintains an official MCP server (@urlbox/screenshot-mcp, a render
tool that returns PNG/PDF/MP4 plus HTML/metadata, env SECRET_KEY). Its differentiator is
forensic-grade output: cryptographic timestamps and hashes, change detection, and many formats. There's
no free tier — a 7-day trial and a $100/10K refund guarantee instead — and entry pricing is $9.50/1,000
($19/mo for 2,000). Best for: agents in legal, compliance, or archival workflows where a defensible,
verifiable capture matters more than price.

Microlink

Microlink offers an official MCP integration over a "browser as API" with full
Puppeteer/Playwright control and global edge caching. Its volume pricing is the cheapest full-featured
option here — $49/mo for 46,000 requests (~$1.07/1,000) — with adblock and cookie-banner handling
available even on the free 50-req/day tier. Best for: high-volume agents that want headless-browser
control and the lowest full-featured price at scale.

CaptureKit

CaptureKit ships an official MCP server and pairs screenshots with a Page
Content API (HTML/Markdown/AI summaries) and stealth proxies, from $7/mo for 1,000 credits. That content
side is handy when your agent needs both a screenshot and clean page text to reason over. Best for:
indie developers building agents that need clean web content for an LLM alongside captures.

ApiFlash, ScreenshotAPI.net, Thum.io — no official MCP

ApiFlash is fast and cheap (Chrome on AWS Lambda, $7/mo for 1,000) but has only
third-party MCP wrappers — fine for a quick experiment, riskier for production. ScreenshotAPI.net
has a broad parameter set (element capture, video, PDF, data extraction) from $9/mo but no MCP server at
all. Thum.io is the cheapest raw price ($0.20/1,000) with a no-signup branded free
tier, but full-page is a paid feature, there's no ad/cookie removal or geo proxy, and no MCP. With these
three you'd hand-roll your own tool wrapper around the HTTP API for an agent.

How to choose for an AI agent

• Want the lowest-commitment official MCP with real-browser quality? Site-Shot — $5/mo, official site-shot-mcp, ad/cookie removal to keep vision tokens down, and a free browser tool to test first.
• Want the richest, most integration-heavy agent tooling regardless of price? ScreenshotOne.
• Need verifiable, forensic-grade captures for compliance or archival agents? Urlbox.
• High volume and want the cheapest full-featured option? Microlink at scale.
• Need clean page content plus screenshots for your LLM? CaptureKit.
• No official MCP, but cheap or simple is the priority? ApiFlash (fast/budget, 3rd-party wrapper), ScreenshotAPI.net (broad params), or Thum.io (cheapest thumbnails) — expect to write the tool wrapper yourself.

For more detail, see our 8-tool screenshot API comparison,
the cheapest screenshot API guide, and
full-page vs. viewport screenshots
when deciding which tool your agent should call.

FAQ

What is a screenshot API for AI agents?
It's a hosted service that renders a web page in a real browser and returns an image (PNG/JPEG), which an
AI agent calls to "see" a page — for vision-model reasoning, web automation, or visual QA. The
agent-friendly way to expose it is an MCP server, so the model can invoke a screenshot tool directly
instead of you writing custom HTTP glue.

What is a screenshot MCP server?
An MCP server that exposes screenshot capture as one or more tools an AI agent can call. Site-Shot's
official server (npx -y site-shot-mcp, registry id io.github.site-shot/site-shot-mcp) exposes
capture_screenshot and capture_full_page; ScreenshotOne, Urlbox, Microlink, and CaptureKit ship
official servers too. ApiFlash has only third-party wrappers; ScreenshotAPI.net and Thum.io have none.

How do I take a website screenshot programmatically for an agent?
Two ways. Over MCP: add the vendor's MCP server to your client config (for Site-Shot, the npx/args/
env block above with your SITESHOT_API_KEY) and let the agent call the tool. Or directly: send a
single HTTP GET to the screenshot API with the target URL and options, and read back the PNG/JPEG or a
JSON payload. Site-Shot supports both; the HTTP API is a plain GET with no SDK required.

Do I need an API key to use a screenshot MCP server?
Yes. The MCP server authenticates with an API key (SITESHOT_API_KEY for Site-Shot), which requires an
account. Site-Shot's free, no-signup option is its browser-based online tool only — that tool does not
issue an API key. The paid API, which the MCP server uses, starts at $5/mo for 2,000 screenshots.

Why remove ads and cookie banners before sending a screenshot to a vision model?
Vision models bill by image tokens, and ads, cookie overlays, and modals add pixels and visual noise you
pay to encode — and they can hide the content you actually want analyzed. Removing them at capture time
(Site-Shot's no_ads and no_cookie_popup) yields a cleaner, often smaller image: lower token cost per
call and more accurate answers.

Which screenshot API is best for building an AI agent in 2026?
If you want the lowest-commitment official MCP with real-browser quality and ad/cookie removal, Site-Shot
($5/mo, site-shot-mcp). For the richest agent tooling and SDKs, ScreenshotOne. For forensic/compliance
captures, Urlbox. For cheapest at volume, Microlink. For screenshots plus clean page content, CaptureKit.

Capture your first screenshot free in your browser — no signup — at
site-shot.com. Building an agent? See the
AI agents page, grab an API key on the
pricing page, and read the
API docs.