MCP server

Take website screenshots from Claude, ChatGPT, Cursor and any other app that supports the Model Context Protocol (MCP). Ask for a page and a device in plain language; the screenshot appears in the chat in a URLBooth card, and the assistant can see it too.

Last updated

Server URL
https://urlbooth.com/mcp
Transport
Streamable HTTP
Authentication
OAuth 2.1 sign-in, or an API key
Plan
URLBooth Pro, $20/month

Overview

The URLBooth MCP server gives an AI assistant four tools:

  • capture_screenshot captures a public page at the exact viewport of any of 30 phones, tablets, laptops and desktops, optionally as a full-length page or inside device artwork, with cookie banners and ads removed.
  • capture_screenshots takes up to 6 screenshots in one call, for example one page on several devices, and shows them together in one card.
  • get_screenshot shows captures from the last 2 days again.
  • list_devices lists every device preset and its size.

It runs the same capture engine, limits and storage as the REST API. Nothing is installed on your computer: the server is hosted at https://urlbooth.com/mcp.

Before you start

  • A URLBooth account with an active Pro plan. You can connect without one, but screenshots will be refused until you subscribe.
  • An MCP-capable app. Claude supports custom connectors on every plan (Free is limited to one). In ChatGPT, custom apps need developer mode, available on Plus, Pro, Business, Enterprise and Edu.
  • For Claude, ChatGPT, Cursor, VS Code, Windsurf and Codex you sign in with OAuth; no key is needed. For other clients, create an API key under Dashboard → API keys.

Connect your app

Claude (claude.ai, desktop and mobile)

  1. Open Customize → Connectors.
  2. Click + Add, then Add custom connector.
  3. Name it URLBooth and enter the server URL https://urlbooth.com/mcp, then click Continue.
  4. Keep the detected sign-in settings. Under OAuth client, keep automatic registration; there is no client ID or secret to enter. Click Add.
  5. Click Connect, sign in to URLBooth with Google or GitHub and click Allow.
  6. In a chat, open the + menu → Connectors and make sure URLBooth is switched on.

ChatGPT

  1. Open Settings → Apps & Connectors → Advanced settings and turn on Developer mode. On Business, Enterprise and Edu, a workspace admin may need to allow it first.
  2. Back in Apps & Connectors, click Create.
  3. Enter the name URLBooth, a short description such as “Website screenshots”, the URL https://urlbooth.com/mcp and choose OAuth as the authentication.
  4. Confirm, then sign in to URLBooth and click Allow.
  5. In a new chat, open the + menu, choose Developer mode and enable URLBooth for the conversation.

Claude Code

Add the server, then run /mcp in Claude Code, select urlbooth and choose Authenticate. Your browser opens to sign in.

Claude Code with OAuth
claude mcp add --transport http urlbooth https://urlbooth.com/mcp

To use an API key instead (for example on a server without a browser), pass it as a header. Add --scope user to make the server available in every project.

Claude Code with an API key
claude mcp add --transport http urlbooth https://urlbooth.com/mcp \
  --header "Authorization: Bearer $URLBOOTH_API_KEY"

Cursor

Add this to ~/.cursor/mcp.json (all projects) or .cursor/mcp.json (one project). Open Cursor Settings → MCP (Tools & MCP in newer versions) and click Connect next to urlbooth to sign in.

Cursor mcp.json
{
  "mcpServers": {
    "urlbooth": {
      "url": "https://urlbooth.com/mcp"
    }
  }
}

VS Code (GitHub Copilot)

Add this to .vscode/mcp.json, or run MCP: Add Server from the command palette and choose HTTP. VS Code asks you to sign in when the server starts. Use the tools from Copilot Chat in agent mode.

VS Code mcp.json
{
  "servers": {
    "urlbooth": {
      "type": "http",
      "url": "https://urlbooth.com/mcp"
    }
  }
}

Windsurf

Add this to ~/.codeium/windsurf/mcp_config.json, then refresh the MCP servers in the Cascade panel and sign in when asked. If your version cannot sign in to remote servers, add an Authorization header with an API key, as in the generic configuration below.

Windsurf mcp_config.json
{
  "mcpServers": {
    "urlbooth": {
      "serverUrl": "https://urlbooth.com/mcp"
    }
  }
}

Codex CLI

Register the server, then sign in. On a machine without a browser, use an API key with the generic configuration below.

Codex commands
codex mcp add urlbooth --url https://urlbooth.com/mcp
codex mcp login urlbooth

Other MCP clients

Any client that supports remote Streamable HTTP servers works. Clients that support MCP authorization discover sign-in on their own from the server’s 401 response; just give them the URL. Clients that only send fixed headers can use an API key:

Generic MCP configuration with an API key
{
  "mcpServers": {
    "urlbooth": {
      "type": "http",
      "url": "https://urlbooth.com/mcp",
      "headers": { "Authorization": "Bearer sk_…" }
    }
  }
}

Only remote (HTTP) connections are supported; there is no local stdio package. Keep API keys out of shared or committed files.

Example prompts

You do not need to name the tools. Describe what you want to see:

  • “Take a screenshot of stripe.com.”
  • “Show me linear.app on an iPhone 17 Pro, inside the phone frame.”
  • “Capture the full page of https://vercel.com/pricing as a JPEG.”
  • “Screenshot our landing page on desktop, iPad and Pixel and compare the hero sections.”
  • “Capture nytimes.com with cookie banners left in place.”
  • “Which device presets can you capture? List the Android ones.”

When you ask for several screenshots, the assistant takes them in one call (up to 6) and shows them in one card. Each one counts as a screenshot toward your daily limit.

Tools

capture_screenshot

Captures a public page and shows it in the chat. Usually takes 3 to 15 seconds. Parameters are the same as the REST API’s:

ParameterTypeDefaultDescription
urlstringrequiredPublic http or https URL, up to 2,048 characters. A missing scheme defaults to https://, so example.com works. Credentials in the URL and private or local hosts are rejected.
devicedevice iddesktopViewport and frame preset, for example macbook-pro-14, iphone-17-pro or ipad-air-11. Ask the assistant to list devices, or see the device list in the API reference.
formatpng | jpeg | webppngImage format of the response.
fullPagebooleanfalseCapture the whole scrollable page. Very large pages are rejected.
cleanupbooleantrueBest-effort removal of cookie banners, consent pop-ups and ads: ad-network and consent-manager requests are blocked, common reject and close buttons are clicked, and empty ad slots are collapsed. Pages that block automated browsers, and banners inside shadow DOM, are not cleaned.
framebooleanfalseDraw the capture inside the matching device: MacBook, iMac, iPhone, iPad and more.

It returns:

FieldDescription
screenshotsOne entry per requested screenshot, in request order. Each entry has the fields below.
idScreenshot id. Pass it to get_screenshot to show the capture again.
urlThe page that was captured, after normalization.
device / deviceLabelPreset id and its name with the viewport, for example iPhone 17 Pro / 17 / 16 Pro · 402 × 874.
formatpng, jpeg or webp.
fullPage, frame, cleanupThe options that were applied. frame is false when the device has no artwork or fullPage is set.
sizeBytesSize of the image file.
imageUrlSigned HTTPS link to the image, valid until expiresAt. The card loads the picture from it, and the assistant can share it with you.
expiresAtWhen the stored capture and its link stop working (ISO 8601), 2 days after capture.
errorOnly on a failed entry: { code, message }. The other entries in the call are unaffected.

The result also carries the images themselves (up to 4 MB in total) so the model can look at them, and a short text summary.

capture_screenshots

Takes 1 to 6 screenshots in one call. Pass shots, a list of objects with the same parameters as capture_screenshot. Up to 3 are captured at a time, and a failure in one entry does not stop the others; it appears in the card with its reason. Returns the same screenshots list. Chat apps show one card per reply, so this is how the assistant shows several screenshots together.

capture_screenshots arguments
{
  "shots": [
    { "url": "https://apple.com", "device": "iphone-17-pro", "frame": true },
    { "url": "https://apple.com", "fullPage": true, "format": "jpeg" },
    { "url": "https://apple.com", "device": "macbook-pro-14", "frame": true }
  ]
}

get_screenshot

Shows earlier captures again: pass id for one or ids for up to 6. Returns a fresh card and signed links with the same fields. Works for 2 days after capture and only for the account that took it. Read-only.

list_devices

Returns every preset as id, label, group, width and height (CSS pixels). Read-only and does not count toward any limit. The full table is in the API reference.

What you see in the chat

In Claude and ChatGPT, screenshots appear in a URLBooth card showing the site, the device, format and file size, the image, and buttons to Open full size or Download. When a call returns several screenshots, a thumbnail strip and arrows let you switch between them (arrow keys work too), and failed entries show their reason. Full-page captures scroll inside the card, and Expand opens the card full screen where the app supports it. The card follows the app’s light or dark theme.

Apps that do not support MCP Apps cards (most editors and terminals) show the image directly, and the assistant can give you the image link from the result.

Authentication

OAuth (recommended)

When an app connects without credentials, the server replies 401 with a pointer to its protected resource metadata. The app registers itself (dynamic client registration), sends you to URLBooth to sign in with Google or GitHub, and shows a consent page naming the app and where it sends you back. After you click Allow, the app receives a token that works only for https://urlbooth.com/mcp.

  • Authorization code flow with PKCE (S256 only). Access tokens last one hour and refresh tokens rotate on every use.
  • Scope screenshots: take and view screenshots. The app cannot see your billing, API keys or other account settings.
  • Every call re-checks that the app is still connected and that your plan is active.
  • Disconnect an app at any time under Dashboard → API keys → Connected apps. Access ends immediately.

API key

Send Authorization: Bearer sk_… on every request, the same key as the REST API. Revoking the key in the dashboard cuts off the client.

Limits and billing

  • MCP captures share the Pro limits with the REST API: 10,000 screenshots per day (reset at 00:00 UTC), 60 per minute and 10 at the same time.
  • Failed captures, get_screenshot and list_devices are not counted.
  • No extra charge and no per-screenshot fees. Requests over a limit return an error that says when to retry.

Privacy and security

  • Screenshots are stored privately in your account and deleted 2 days after capture.
  • Each result includes a signed image link so the chat app can display the picture. Anyone who has the link can open that one image until it expires with the capture; it gives no access to anything else. Treat it like a share link.
  • The assistant’s app (Anthropic, OpenAI and so on) receives the screenshot and its link as part of your conversation, under that app’s own privacy terms.
  • Only public pages are captured; private, loopback and internal addresses are refused. See the privacy policy.

Troubleshooting

The tools do not appear
Start a new chat after connecting, and check the connector is switched on for the conversation (+ menu). In ChatGPT, developer mode must be selected for the chat. In editors, reload the MCP server list.
The app keeps asking me to sign in
Finish the sign-in in the browser window the app opens and click Allow on the consent page. If you disconnected the app in the dashboard, or signed in with the wrong account, remove the connector and add it again.
I see the text answer but no picture
Older clients may not render cards or images. Ask the assistant for the image link, which opens the picture in your browser. In ChatGPT, update the app and use a new chat: the card is drawn by ChatGPT’s app framework.
The tool says subscription_required
The URLBooth account you signed in with has no active Pro plan. Subscribe, or disconnect and sign in with the account that has the plan.

Tool errors start with a stable code:

CodeWhat to do
subscription_requiredThe signed-in URLBooth account has no active Pro plan. Subscribe on the pricing page, then ask again.
invalid_requestThe URL is not a public http(s) address, or an option is invalid (for example an unknown device id).
capture_failedThe site refused automated browsers, timed out or returned an error page. Try again or try another page.
page_too_largeThe full page is taller than 10,000 px or wider than 4,096 px. Capture without fullPage.
image_too_largeThe image is over 10 MB. Use jpeg or webp, or capture without fullPage.
rate_limitedMore than 60 requests a minute or 10 captures at once. The message says when to retry.
daily_limit_reachedThe 10,000 screenshots for today are used. The limit resets at 00:00 UTC.

Still stuck? Email hi@urlbooth.com with the app you use and the time of the failed request.

FAQ

Does the URLBooth MCP server cost extra?

No. It is included in the Pro plan ($20 per month) and shares its limits with the REST API. There are no per-screenshot charges.

Do I need an API key to use URLBooth in Claude or ChatGPT?

No. Claude and ChatGPT connect with OAuth: you sign in to URLBooth with Google or GitHub and approve access. API keys are only needed for clients that send a fixed header.

Why does ChatGPT ask me to confirm every screenshot?

capture_screenshot stores a new capture in your account and counts toward your daily limit, so it is marked as an action that changes data. ChatGPT asks before running such tools. You can allow it for the rest of the conversation.

Can the assistant see the screenshot?

It depends on how the client connects. With an API key (IDEs, CLI agents, scripts), images up to 4 MB are attached to the tool result so the model can describe, compare or check them. In ChatGPT and Claude, which sign in with OAuth, you see the images in the card and the model gets a signed link to each one: chat apps do not pass a large result on to the card, so the image is not attached there.

How do I disconnect an app?

Open Dashboard → API keys → Connected apps and click Disconnect. Access ends immediately, including any tokens the app already holds.

Which pages can it capture?

Any public http or https page. Pages behind a login, on private networks or on localhost cannot be captured.