Webshot Now Speaks MCP: Screenshots as a Tool for AI Agents

Webshot.site is now capable of understanding the Model Context Protocol (MCP), allowing AI agents to render web pages and view their contents directly without the need for copy-pasting URLs or manual description. To utilize this feature, simply configure your client with the following single line of configuration: {"mcpServers": {"webshot": {"url": "https://webshot.site/mcp"}}}. The endpoint, accessible at https://webshot.site/mcp, supports JSON-RPC 2.0 over HTTP POST.

The service exposes a single tool, capture_sandbox_webshot, which accepts a URL and optional viewport, full_page, and format parameters. To request a screenshot of example.com and ask your assistant to focus on what is above the fold, you would simply need to provide this URL to the capture_sandbox_webshot tool. It will then return a JSON envelope containing a link to the screenshot image, rather than the image itself in base64 format.

This approach prevents the image from consuming too much of the model's context window and potentially causing issues with the conversation.

The returned JSON envelope includes essential information such as the image format (jpg), the mode (desktop_viewport), the size of the image in bytes (26,600), the number of credits consumed (1), and a 128-bit random token for authorization. This token is unguessable and serves as the only way to access the image. Unlike the image itself, the link is public, session-free, and reusable, allowing you to share it with other tools, include it in tickets, or access it directly.

Notably, you don't need an agent to utilize this feature. The same JSON envelope is available through plain HTTP by adding &response=json to a normal capture request. This allows you to retrieve the raw image data if needed. The existing API remains unchanged, with the addition of this new functionality. The credits model is also worth understanding, as the allowance is 5 credits per 15 minutes per IP, and credits are not tied to the number of captures.

A default anonymous call consumes 2 credits, and you receive 5 captures per window. Switching to full_page=false or mode=desktop_viewport reduces the cost to 1 credit and increases the number of captures to 5.

The service is inherently slow, as it performs a real headless-Chrome render, loading the page, waiting for it to settle, and capturing it. This process takes 10-60 seconds. Ensure that your MCP client timeout is set to 90 seconds or more to avoid spurious failures due to timeouts. Additionally, in-browser agents can leverage the same tool through the document.modelContext API, which is part of the emerging W3C browser standard for exposing page capabilities to AI.

This works natively in Edge 147 and Chrome 149 during their origin trials, using your existing session and quota without requiring an API key in the model's context.

Safety is a priority with Webshot.site. The capture runs in an isolated sandbox, and it refuses access to private, loopback, and cloud-metadata addresses. Pointing the service at a suspicious URL serves as the intended use case, allowing you to inspect a page's appearance without directly loading it in your own browser. Machine-readable documentation for agents is available at /llms.txt, and the full human reference, including the MCP section, can be found on the developers page.