Site-Shot
FOR AI AGENTS

Screenshots for AI agents

Give Claude, Cursor, and your AI agents the ability to see any web page. Site-Shot's website screenshot API plugs into your agent stack over MCP, GPT Actions, or a plain HTTP call — real Chromium rendering, full-page capture, country proxies, and automatic ad & cookie-banner removal so the model reads the page, not the overlay.


Install the MCP server

You need a Site-Shot API key; captures draw on that account's own API allowance. A confirmed account's key comes with the free plan — 100 captures a month, no card — and a paid plan adds volume, geotargeting, storage and schedules. Get your key on the quickstart (it is also in the dashboard), and pricing lists the plans. The browser capture tool is free and needs no account, but it issues no API key.

Claude Code and Codex CLI can add the server from the command line; both commands fetch the published site-shot-mcp package from npm:

# Claude Code: --scope user adds it for every project, not just this directory;
# export SITESHOT_API_KEY first so the key stays out of your shell history
claude mcp add --scope user site-shot -e SITESHOT_API_KEY="$SITESHOT_API_KEY" -- npx -y site-shot-mcp

# Codex CLI: add the entry, then put env_vars = ["SITESHOT_API_KEY"] under
# [mcp_servers.site-shot] in ~/.codex/config.toml and export the key in your shell
codex mcp add site-shot -- npx -y site-shot-mcp

Or install it as a plugin. In Claude Code the plugin takes the key in its own settings rather than on the command line:

claude plugin marketplace add site-shot/site-shot-mcp
claude plugin install site-shot@site-shot
# then, inside Claude Code, enter the key:
/plugin configure site-shot@site-shot

Until the key is set, Claude Code does not start the server and the capture tools do not appear. The setup guide for each client:

Other clients that support local stdio MCP servers — Claude Desktop, Cursor, Cline — can use this manual configuration:

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

One route per client: the command line, the plugin and this manual entry are separate configurations of the same server, not steps.

Then ask your agent to “take a full-page screenshot of example.com” — it gets the image back inline. Two tools: capture_screenshot and capture_full_page.


Why Site-Shot for agents

  • Cleaner input for the model. Automatic ad + cookie-banner removal (no_ads, no_cookie_popup) means the screenshot shows the page, not a consent dialog. It does not shrink the token bill — vision models charge by image dimensions — so choose the viewport for that.
  • Real Chromium. Pages render exactly as a real visitor sees them — JS, fonts, lazy-loaded images.
  • Full-page capture up to 20,000px, or precise viewport/device sizing.
  • See the world's web. Country-specific proxies set IP, language, time zone, and geolocation.
  • One call, image back. No browser cluster to run, no SDK required — or install an official SDK (npm install site-shot-sdk, pip install site-shot) if you're writing JavaScript or Python.

Five ways to connect

MethodForHow
MCP serverClaude Code, Codex CLI, Claude Desktop, Cursor, Cline, LangChain, CrewAIPlugin or npx -y site-shot-mcp
Official SDKsNode.js / TypeScript and Python agents and scriptsnpm install site-shot-sdk · pip install site-shot
LangChain toolLangChain agents in Pythonpip install langchain-siteshot
GPT Actions / OpenAPICustom GPTs, OpenAPI clients/openapi.json
Plain HTTP APIAny language / frameworkSee the API schema (example below)

Works with: Claude Code · Codex CLI · Claude Desktop · Cursor · Cline · LangChain · CrewAI · OpenAI GPTs · n8n · any client that runs a local stdio MCP server.

Guides: AI agent vs. screenshot API — who should capture the page · Best screenshot MCP servers in 2026 · Screenshots for AI agents over MCP · Website screenshots for a LangChain agent · 8-tool screenshot API comparison.


Discovery entry points

Use these URLs to discover and integrate Site-Shot from AI agents and agent runtimes.


Recommended agent flow

  1. Read /llms.txt and the relevant Markdown docs.
  2. Import /openapi.json into GPT Actions or another OpenAPI-compatible agent runtime.
  3. List public plans, start email-gated signup, and let the user confirm setup on Site-Shot.
  4. To pay, send the account owner to /pricing/: they choose a plan and confirm payment in Stripe, and the same API key switches on. Only Site-Shot’s own ChatGPT Actions client can open Stripe Checkout itself, with the billing:checkout scope.
  5. Read usage/subscription status and reveal the screenshot API key only through the explicit API-key action.

Security boundaries

Agents do not collect card details. Payments are confirmed by the user in Stripe Checkout. The full screenshot API key is never returned in generic profile or status calls; it requires the explicit api_key:read scoped action.


Quick start

Once a screenshot API key is revealed, capture a full-page screenshot with a single request. The capture API is one GET endpoint — https://api.site-shot.com/ — and it authenticates with the userkey query parameter, not an Authorization header. Every parameter is listed in /openapi-capture.json.

curl -G --fail 'https://api.site-shot.com/' \
  --data-urlencode 'url=https://example.com/' \
  --data-urlencode "userkey=$SITESHOT_API_KEY" \
  --data-urlencode 'width=1280' \
  --data-urlencode 'full_size=1' \
  --data-urlencode 'format=png' \
  --data-urlencode 'response_type=image' \
  --output screenshot.png

FAQ

Is there a screenshot MCP server?

Yes — Site-Shot ships an open-source MCP server (site-shot-mcp) that gives Claude and other AI agents two tools to capture website screenshots, with real Chromium rendering and automatic ad/cookie-banner removal.

How do I give Claude the ability to take screenshots?

Add the Site-Shot MCP server to your Claude Desktop config (snippet above) with your API key. Claude can then capture any URL and view the image inline.

Why use a screenshot API instead of Puppeteer for my agent?

You skip running and scaling a headless-browser cluster, and you get proxies, ad/cookie removal, and full-page capture out of the box — a single call returns the image. The full architecture argument — when the agent should capture and when it should delegate — is in AI agent vs. screenshot API.

Does it reduce token usage for vision models?

Not by itself. Vision models charge by image dimensions, not by what the pixels show, so a 1920×1080 capture costs the same tokens with or without a banner in it. Removing ads and cookie overlays gives the model a cleaner image to read; to spend fewer tokens, capture a smaller viewport or crop to the region you need.