Websites Directly Inform AI Agents How to Use Them

In the past, the web evolved to be optimized for human visual interaction. However, the ability of agents to understand and control websites will become increasingly important in the future.

In line with this change, there is a forecast that the survival and success of web services may depend on designing agent-friendly websites.

This technology provides a standardized way to directly inform AI about how to use a website, helping AI agents utilize the web more efficiently.

Existing browser agents operate by analyzing screenshots to use websites, leading to excessive token consumption. Additionally, there are limitations such as significantly slower task speeds due to repeated processes of understanding web structures and analyzing tools.

This method also involves the inconvenience of analyzing the DOM (Document Object Model) and accessibility trees.

This is possible by pre-defining the website's structure at the code level.

As a result, AI can directly call the necessary tools to perform desired tasks without having to parse the website's structure.

This declarative method bypasses the complex guessing processes of existing agents and enables efficient interaction.

In the airline booking website demo, a total of four WebMCP tools were implemented. These tools include 'ListFlights' to check flight lists, 'ResetFilters' to reset filters, 'SetFilter' to set search filters, and a tool to search for flights.

It was confirmed through the demonstration that the agent automatically handles all processes from searching to filtering with a single prompt.

Existing MCP (Machine Comprehensible Pages) required requests through a separate server for agents to interact with websites. However, a major difference with WebMCP is the absence of an intermediary server.

When a web page is opened, WebMCP tools are already defined within that page, allowing agents to directly utilize the website's inherent logic and screen.

As of August 2026, several coding agent products are gradually adopting this technology, marking its early introduction phase.

Currently, it is an early entry point with not many related services, and it is expected to be applied to more services in the future.

When WebMCP was applied to a blog, the agent could recognize and utilize the site's tools.

This demonstrates the efficiency of searching and filtering blog posts by the agent through input schema definitions, compared to existing browser agent methods.

The existing method using browser agents required multiple steps such as understanding page structure, taking screenshots, and clicking specific buttons. In contrast, registering WebMCP tools eliminates these complex steps.

Agents can directly call the registered WebMCP tools to perform tasks, thereby saving token usage and efficiently reducing task processing time.

Previously, Search Engine Optimization (SEO) was crucial when developing websites, but with the advent of AI services like Claude, Answer Engine Optimization (AEO) has emerged as a new area of interest.

The next step in this evolution is to use WebMCP to enable agents to directly run browsers and complete tasks within websites. This shift to an agent-centric web environment presents a new paradigm in web development.

The declarative API is implemented by specifying tool names and descriptions within HTML forms.

The imperative API, on the other hand, uses JavaScript and JSON schema to register and call tools. Both methods support integrating WebMCP functionality into existing websites.