We made web-works.io one of the first agency sites AI agents can operate (not just read)

We just did something new on our website. If you visit web-works.io with an AI agent in Chrome 149, that agent can now run our site. It does not just scrape our pages or guess where to click. It uses clean, structured tools to find information and fill out forms.

We built this using WebMCP, a new protocol from Google and Microsoft. We are one of the first web agencies to put this technology into practice. Why did we do it? We wanted to prove we can build the future, not just talk about it.

What is WebMCP and why does it matter?

Most AI agents browse the web like humans. They look at screenshots, read the text, and try to click the right buttons. This approach is slow. It often fails when a design changes. It also uses a lot of computer power. Today, if you ask an AI assistant to buy shoes on a website, it has to look at the screen. It takes a screenshot, analyzes the pixels, and guesses where to click. If a pop-up appears, the AI gets confused. If a button color changes, the AI might miss it. This process is incredibly slow. It also costs a lot of money because sending screenshots to AI models uses massive amounts of data.

The Web Model Context Protocol, or WebMCP, changes this. It is a proposed browser-native standard. It lets a website expose structured tools directly to any AI agent running in the browser. The W3C Web Machine Learning Community Group is incubating this specification. The draft was first published on February 10, 2026. Its co-editors include Brandon Walderman from Microsoft, alongside Khushal Sagar and Dominic Farolino from Google.

Khushal Sagar, a Staff Software Engineer for Chrome at Google, described WebMCP as:

“the USB-C of AI agent interactions with the web.”

Instead of guessing, the agent gets a list of exact functions it can call. This makes a massive difference in performance. Early benchmarks show that WebMCP achieves about 98% task accuracy. Vision-based browsing only gets about 45% accuracy. WebMCP also reduces token use by up to 89% and cuts computational overhead by 67%. It is faster, cheaper, and far more reliable.

How we set up WebMCP on our site

We wanted to show our clients how this works in the real world. We did not want to just write about it. We wanted to build it.

Our team added WebMCP support to our own site in Chrome 149. The entire setup took about 200 lines of code. It has zero impact on our human visitors. If you browse our site normally, nothing looks or acts any differently.

We used the protocol to expose four specific tools to AI agents:

• List our services: The agent can ask for our core offerings instantly.
• Search our blog: The agent can find articles on web design, SEO, or performance.
• Fill in the contact form: The agent can pre-fill our contact fields based on your request.
• Switch the site language: The agent can change the interface language for you.

Let us look at how developers actually build this. WebMCP gives us two ways to connect our site to an agent. The first is the imperative API. With this method, we write JavaScript functions to handle complex tasks. We register these functions using navigator.modelContext.registerTool(). We define the exact inputs the agent must provide using a JSON Schema.

The second way is the declarative API. This is even simpler. We do not have to write complex JavaScript. Instead, we can take our standard HTML forms and add simple attributes. We use attributes like toolname and tooldescription directly on the form elements. The browser reads these attributes and automatically shows them to the agent as tools. This means any simple form can become AI-ready in minutes. You can read more about these methods on the Chrome for Developers portal.

Keeping humans in control

We made one crucial design decision during this project. An AI agent can fill out our contact form, but it cannot submit it. A human must always review the form and click the submit button.

This is not a technical limitation. We designed it this way on purpose. For our German-speaking clients, data privacy is a major concern. Keeping a human in the loop ensures we respect GDPR consent. The human user always stays in control of their own data. AI should help you do the boring work, but it should not speak for you without your permission.

The agentic web is arriving

Why did we spend time on this? The web is changing fast. Soon, many users will not browse websites themselves. They will send AI agents to do research, buy products, or book services for them.

Websites will need more than just pretty pixels for human eyes. They will need clear APIs for AI agents. Agencies that only design for human screens will leave their clients behind.

We wanted to prove that we can build these systems today. Google launched an official WebMCP origin trial in Chrome 149. We jumped on it immediately. If the standard rolls out to everyone later, we will already have the experience. We did not wait for the future. We built it in 2026.

If you want your website to be ready for the next generation of web users, we can help. Let us know what you want to build. We are ready.