Web Model Context Protocol

Programs like ChatGPT or Gemini can now not only answer, but also carry out tasks. They are meant to, for example, search for a flight or fill a shopping cart. So far they often do this by remote-controlling a browser and clicking buttons like a human would. The Web Model Context Protocol reverses this relationship: the website itself describes what can be done on it. It offers the assistant a list of clearly named actions, for example “search product” or “book appointment.” The assistant then no longer has to guess where which button is located.

Why clicking by bots works so poorly

When an AI operates a website, it basically only sees a pile of text and pixels. It has to infer from that which element is a search field and which is an advertising banner. This often works, but not reliably. If the operator changes the layout, the process suddenly fails. For companies this is unusable, because an ordering process must not break off midway in ten percent of cases.

On top of that comes the effort involved. An assistant that analyzes screenshots of a page consumes a great deal of computing time, and computing time costs money. A direct call to a described function is many times cheaper and faster. That is why many providers see this as the prerequisite for AI assistants to be able to shop or book on a large scale.

It is also about control. A shop wants to determine for itself what an outside program is allowed to do on its site. Displaying prices: sure. Triggering a hundred orders per second: rather not. If the site itself defines the permitted actions, the operator retains control over it. It is precisely this promise that makes the proposal interesting for companies.

How a site offers its actions

The basic idea resembles a menu. A restaurant does not let guests rummage around in the kitchen, but instead hands over a list of what can be ordered. In the same way, a website deposits a list of its functions. Each function has a name, a short description in plain language, and information about what data it requires.

Technically, the site registers these functions in the browser, usually via a small amount of JavaScript code. JavaScript is the programming language that runs alongside almost every website. An AI assistant working in the same browser can query the registered functions and call one of them. The site then carries out the action using its own code and returns a result.

It is important here that the user can already be logged in. The assistant works within the human’s active session and does not need any passwords. Sensitive steps such as a payment can be designed so that the human must explicitly confirm them. The name of the procedure refers to the Model Context Protocol, an older standard with which AI models use tools on servers. The web counterpart transfers this idea to ordinary websites.

State of standardization and early applications

The proposal dates from 2025 and is being driven forward by, among others, Microsoft and Google. It is being discussed in a working group of the W3C, the organization that sets the rules for the web. It is not yet a finished standard. Anyone reading about it today is reading about drafts, test versions, and individual browser extensions.

Nevertheless, the term regularly appears in business news. The reason is trade on the internet. If customers in future send an assistant out to shop, this technology will help decide which shop is even found in the first place. Analysts are therefore discussing whether search engine advertising is losing importance. The approach should not be confused with a classic programming interface: that operates behind the browser, whereas here the visible page itself remains the point of access.