Webshot adds MCP endpoint so AI agents can screenshot and inspect web pages

What launched

Webshot.site, a service that renders web pages in a headless browser and returns screenshots, has added support for the Model Context Protocol, the standard that clients like Claude Desktop and Cursor use to expose capabilities to AI agents. According to a post on dev.to, any MCP-capable client can now point at the service's /mcp endpoint and gain the ability to capture a page as a tool call, with no API key or signup required for basic use.

The endpoint speaks JSON-RPC 2.0 over HTTP POST and exposes a single tool, capture_sandbox_webshot, which takes a URL plus optional viewport, full_page and format parameters. Once configured, a user can ask an assistant to screenshot a site and describe what sits above the fold; the agent invokes the tool, receives the result, and answers from it.

A link instead of raw bytes

The most deliberate design choice is what the tool returns. The obvious implementation would base64-encode the PNG and push it into the model's context, but the developer notes that a typical screenshot runs 25–200 KB, which becomes a wall of base64 that crowds out the conversation while most agents cannot do anything useful with those bytes anyway.

Instead, the tool replies with a compact JSON envelope containing a hosted image URL, the output format, the capture mode, the byte count and the credit cost. The link is public, session-free and reusable, so it can be handed to another tool, pasted into a ticket, or opened by a human. Access is guarded only by a 128-bit random token, which the post describes as unguessable.

The same envelope is available outside agent workflows: adding a response= parameter to the service's regular capture endpoint returns the JSON form, while requests without it behave exactly as before. The change is strictly additive.

Credits, throttling and latency

Free, keyless usage is metered in credits rather than captures, and the post calls out this distinction as the detail people get wrong when planning. The allowance is five credits per 15-minute window per IP, but a viewport capture costs one credit while a full-page or custom-size capture costs two. Because the default mode is full-page, an anonymous default call spends two credits, meaning two calls per window rather than five. Setting full_page to false, or requesting viewport mode over plain HTTP, restores the one-credit cost, and the output format does not affect pricing. A throttle-status endpoint reports the live balance, and paid plans start at $1.99 per month with keys issued immediately at checkout.

Latency is the other operational caveat. Each capture is a real headless-Chrome render that loads the page, waits for it to settle, and photographs it, taking 10 to 60 seconds. The recommendation is to set MCP client timeouts to 90 seconds or more to avoid spurious failures on captures that are actually fine.

In-browser agents and sandboxing

Beyond MCP, pages on the site register the same tool through document.modelContext, an emerging W3C browser API for exposing page capabilities to AI. It works natively in Edge 147 and behind an origin trial in Chrome 149. Because those calls are same-origin, they draw on the visitor's existing session and quota, which means no API key ever enters the model's context — a useful property when handing tools to something that might quote secrets back.

On safety, captures run in an isolated sandbox, and requests aimed at private, loopback or cloud-metadata addresses are refused. That makes the service deliberately suited to inspecting suspicious URLs: an agent can see what a page looks like without your own browser going there. Machine-readable documentation for agents lives in an llms.txt file, alongside the full human reference on the developers page.

Why it matters

Screenshots are among the simplest ways to give a text-only agent ground truth about a visual, layout-dependent web, and MCP is becoming the wiring standard for exactly this kind of capability. Webshot's implementation demonstrates two patterns worth copying elsewhere: return hosted links rather than base64 blobs to protect the context window, and keep credentials out of the model entirely where same-origin browser APIs make it possible. The credit model and 10-to-60-second render times are real constraints for high-volume scripting, but for ad-hoc agent inspection of pages — including dubious ones — a keyless, single-tool integration is a cheap addition to any agent toolbox.