Give your AI assistant eyes:
website screenshots through MCP

"Does our pricing page still show the old plan names?" "What does the checkout look like on an iPhone?" "Summarise what this competitor changed on their home page." These are questions about how a page looks now, and a language model on its own cannot answer them. It was trained on the web as it was, and a plain HTTP fetch misses everything a page builds with JavaScript, every cookie banner in the way, and what the page actually looks like.

The fix is not a bigger model. It is a tool: a way for the assistant to ask a real browser to load the page and hand back an image, a PDF or the page's text. That is what an MCP screenshot server provides.

The Model Context Protocol is an open standard for connecting AI assistants to tools and data. A server describes the tools it offers, such as "capture this URL", along with their options. A client such as Claude, Cursor or VS Code shows those tools to the model, and the model decides when to call one while it works on your request. You add a server once; from then on the assistant can use it in any conversation.

• capture_screenshot captures a URL as an image: full page or viewport, size, device, ad and pop-up blocking, elements to hide, and the country to load it from.
• capture_pdf renders a URL as a PDF, with selectable text or as an image. A long PDF can outlast the call; the assistant then gets the capture id and fetches it with get_capture once it is ready.
• read_page returns the page's main content as markdown or plain text, without navigation, ads, cookie banners or scripts, for when the assistant needs what a page says rather than how it looks. The page is loaded in a real browser, so pages built with JavaScript work.
• list_capture_options lists the options and their defaults, so the assistant builds valid requests rather than guessing.

Each capture comes back with its details and a download link. A small image is also returned inline so the assistant can look at it; a large one stays a link rather than filling the conversation.

Sign in (no key to copy)

Add the address as a custom connector in your assistant. If it asks for an OAuth client ID, enter Mcp and leave the secret empty. You are sent to our sign-in page, you approve the access, and on the first capture the server creates an Instant Snap API key for itself, named MCP connector with the date, in your workspace. The person signing in needs a role that can create API keys.

An API key in a header

Create an Instant Snap API key (it starts with wsw_api_) and have the assistant send it on every request as X-Api-Key. In Claude Code that is one command:

claude mcp add --transport http wsw https://mcp.websitescreenshotworld.com/mcp --header "X-Api-Key: wsw_api_..."

Other assistants take a short JSON entry with the same address and header. The MCP server documentation has the exact configuration, including the bridge for clients that only speak stdio.

• Review a release. "Capture our pricing page full page on desktop and on an iPhone, and list anything that looks broken on mobile."
• Check another market. "Capture this product page from Germany and from the United States and compare the prices and currency shown." Country selection depends on your plan; the assistant can check what is available.
• Read before you summarise. "Read this changelog page and give me the three changes that affect our integration." read_page fetches the current text instead of relying on what the model remembers.
• Make an evidence copy. "Save this terms page as a PDF with selectable text."
• Debug a layout. In a coding assistant: "Capture the staging URL after my change and tell me if the header still overlaps the hero."

• Give the assistant its own key. Do not reuse one from production code. Its calls are then easy to tell apart in the Instant Snap API history, and you can revoke it on its own.
• Metering is the API's. Each capture counts against your plan and its API rate limit exactly as a direct API call would, because that is what it is.
• Disconnect in the right order. Remove the connector from the assistant first, then delete the MCP connector key. Deleting the key alone is not enough while the assistant is still signed in: the next capture fails, and the one after it creates a new key.
• Treat page content as content. A page the assistant reads is written by whoever owns it. Ask for summaries and comparisons, and be wary of any page text that tries to tell the assistant what to do.

The server covers on-demand captures: the Instant Snap API, which needs a plan that includes it. It does not create or run Auto Schedules. The Auto Schedule API is a single trigger route and is simpler to call directly from your own code or a deploy hook.

An assistant is also the wrong tool for a record you need every day. For monitoring and archiving, set up a scheduled capture and let the assistant work with the results.

An MCP screenshot server lets an AI assistant load a live page in a real browser and come back with an image, a PDF or the page's text. Connect it by signing in or with its own API key, ask for specific pages and variants, and keep scheduled captures for the records you need every day.