WebMCP: when the web moves from showing information to doing work

Until now, the internet has mainly been built for two kinds of visitors: people who click through pages and search engines that read and index content.

A third kind of visitor is now emerging: the AI agent.

An AI agent does not simply read a page and summarise it. It tries to help the user achieve a goal: compare options, analyse a need, prepare a brief, calculate an opportunity or get the next step ready.

The problem is that most websites were not designed for this. An agent often has to interpret menus, buttons and forms much like a person would, without always knowing what is important, safe or permitted.

This is where WebMCP comes in.

What is WebMCP?

WebMCP is a proposed web standard that lets websites expose clear, structured tools to AI agents.[1][4]

Think of it this way.

A conventional website says:

Here are our pages, texts, buttons and forms. Try to work out how everything fits together.

A WebMCP-ready website can instead say:

Here is what I can help with. Here is the information I need. Here is the result you will receive. And here is where human approval is required.

It is the difference between giving someone a picture of a calculator and giving them a working calculator.

Google describes WebMCP as a way to give AI agents faster, more reliable and more precise paths for interacting with a website than trying to operate its visual interface.[2]

The technology supports both straightforward actions, such as clearly defined forms, and more dynamic functions that use the website's own logic.[2] Visitors do not need to understand the technical distinction. What matters is that the agent receives a clear contract for what it can do.

A simple example

Imagine that you run a consultancy and want to understand where AI could have the greatest effect.

On a conventional website, you may need to:

1. Read several service pages.
2. Decide which service seems relevant.
3. Estimate the time and potential value yourself.
4. describe the problem in a contact form.
5. Book a meeting and explain everything again.

On a WebMCP-ready website, your agent could instead:

1. Describe the company's situation in a structured way.
2. Ask the website to assess AI readiness.
3. Match the need to a relevant service.
4. Estimate the automation opportunity.
5. Create an initial workflow brief.
6. Prepare an audit request or booking handoff for your review.

The human remains in control. The agent can prepare the work, while sensitive or binding steps can require a visible click and explicit approval.

That does not merely make the experience more convenient. It changes the role of the website itself.

From digital brochure to active service

Most websites still operate as digital brochures. They display information and hope that the visitor finds the right path.

With WebMCP, a website can become an active part of the customer journey.

It can:

• understand a clearly stated need,
• ask the right questions,
• use its own verified logic,
• present a structured result,
• disclose assumptions and limitations,
• and prepare the next step without taking control away from the user.

This means that tomorrow's best websites may not be judged only by how attractive, fast and searchable they are. They may also be judged by how useful they are to AI agents.

How WebMCP could change the internet

WebMCP is still experimental. Chrome describes it as a proposed standard and currently offers it for testing, prototyping and time-limited origin trials.[1][3] It is too early to know exactly how widely it will be adopted.

But the direction points to several major changes.

1. Websites become actionable

Today's internet is full of information. The next internet will increasingly be designed for action.

A restaurant website could do more than show a menu; it could offer a verified tool for finding dishes based on allergies and the preferences of a group. An insurer could let an agent compare options using approved rules. A software company could help an agent configure the right package without forcing it to guess its way through the interface.

2. AI agents no longer have to pretend to be people

Many agent systems currently work by visually reading a page, locating buttons and simulating clicks.

That can work, but it is fragile. A changed label or relocated menu can make the flow unreliable.

WebMCP lets the website describe its functions directly. An agent can understand that a tool is called calculate_automation_opportunity, what information it requires and what kind of result it returns, without interpreting the design pixel by pixel.

3. A brand's own logic becomes more important

As AI agents mediate more digital interactions, websites risk being reduced to raw material for external models.

WebMCP gives companies an opportunity to define how their services should be used, what rules apply and when a person must take over.

That could make the company's own website more important, not less. It becomes the authoritative place where an agent obtains both information and capability.

4. Conversion takes on a new meaning

A conversion no longer has to begin with a person landing on the home page and clicking a button.

It might begin when a user tells an agent:

Find the best way to automate our lead handling, estimate the potential and prepare a meeting brief.

The agent can then collaborate with the website and return a well-structured decision document. By the time the person enters the process, they are better informed and closer to a decision.

5. Trust and control become competitive advantages

An agent-ready website must not mean that an agent can do anything it wants.

Clear boundaries become more important:

• Which functions are advisory only?
• Which results depend on assumptions?
• When is human review required?
• When is information sent elsewhere?
• Which information should never be given to a tool?

Companies that answer these questions transparently can earn more trust than those that simply automate as much as possible.

WebMCP does not replace the conventional website

WebMCP does not need to replace the interface built for people.

A website should still work in an ordinary browser and remain accessible, fast and easy to understand. WebMCP adds another layer, much like a business offering both a reception desk for people and a clearly organised service desk for digital assistants.

Nor is WebMCP a documented shortcut to better SEO or guaranteed visibility in AI search. Its primary value is making a website's functions more useful and reliable for agents.

How to test WebMCP on patrickpetcu.com

I have implemented WebMCP on patrickpetcu.com as a progressive layer on top of the conventional website.

The site offers seven tools that a compatible agent can use:

• discover the site's agent capabilities,
• assess a company's AI readiness,
• find a relevant AI service,
• create an initial workflow brief,
• calculate automation potential,
• prepare an AI audit request,
• and prepare a booking handoff.

The final two tools deliberately keep a person in control. An agent may prepare a draft, but it cannot submit the contact form or confirm a meeting on your behalf. You review and approve those steps yourself.

Step 1: enable WebMCP in Chrome

WebMCP remains experimental. Chrome documents the following method for local testing:[1]

1. Use a current compatible version of Chrome, Chrome Dev or Chrome Canary.
2. Enter chrome://flags/#enable-webmcp-testing in the address bar.
3. Set the feature to Enabled.
4. Relaunch Chrome.

If the flag is missing, you may need a newer Dev or Canary build. Chrome also lists an origin trial from Chrome 149 for broader testing on live websites.[1][3]

Step 2: install the inspector

Install Google's Model Context Tool Inspector.

Chrome recommends the extension for viewing the tools registered by a page, invoking them manually, inspecting structured output and testing natural-language instructions.[1]

Step 3: open the website

Visit:

https://www.patrickpetcu.com/en

Then open Model Context Tool Inspector. If everything is working, the site's registered WebMCP tools should appear.

Step 4: try ordinary questions

For example:

We are a consultancy with 20 employees. Five people spend four hours a week on manual lead follow-up. We use AI occasionally and our biggest obstacle is unclear processes. Assess our AI readiness and recommend a next step.

Or:

Calculate the automation opportunity if six people save three hours per week, the hourly cost is SEK 650 and 60 percent of the work can be automated.

Or:

Create an initial workflow brief for qualifying and enriching incoming leads before handing them to a salesperson for the final decision.

You can also ask the agent to prepare an AI audit. A visible brief then appears on the website. Nothing is sent until you choose to complete and submit the form yourself.

Alternatively: test the public MCP server

External agent tools can use the public read-only MCP endpoint at:

https://www.patrickpetcu.com/mcp

It provides analysis and information but does not create leads or book meetings. More technical information is available at patrickpetcu.com/developers.

This is only the beginning

WebMCP will not change the internet overnight. Standards take time, browser support changes, and both security practices and user behaviour need to develop.

But the underlying idea is powerful:

Websites should not merely be readable by AI. They should be able to collaborate with AI on the website's own clearly defined terms.

When that happens, the web moves from being a network of documents to also becoming a network of useful capabilities.

The most important question for a company will no longer be only:

What should people be able to read on our website?

It will also be:

What should an AI agent safely be able to help our customers accomplish?

I believe that question will shape the next generation of the internet.

Sources

[1] Chrome for Developers: WebMCP – https://developer.chrome.com/docs/ai/webmcp

[2] Chrome for Developers: WebMCP is available for early preview – https://developer.chrome.com/blog/webmcp-epp

[3] Chrome for Developers: Join the WebMCP origin trial – https://developer.chrome.com/blog/ai-webmcp-origin-trial

[4] WebMCP specification – https://webmachinelearning.github.io/webmcp/