Back to Blog

WebMCP: The Web Standard That Makes Every Website a Tool for Agents

Z

Zak

11th March 2026

AI agent interacting with web applications and network infrastructureAI agent interacting with web applications and network infrastructure

There is a version of AI that most people are already familiar with. You ask it a question, it gives you an answer. You describe a problem, it writes you some text. It is, at its core, a very sophisticated text machine.

That version of AI is about to look extremely outdated.

The Wall AI Has Always Hit: Why Text-Only AI Has a Hard Limit

Until very recently, there were only two ways for an AI agent to interact with a website or web-based system.

The first is visual. The agent looks at screenshots, reads the text on screen, and tries to figure out where to click, just like a person squinting at an interface they have never used before.

The second is semantic. The agent parses the raw underlying structure of the page, the accessibility tree, the DOM, trying to extract data and trigger actions buried in thousands of lines of code it was never meant to read directly.

Both approaches are slow, expensive, and unreliable. Every action requires the AI to reason its way through something that was built for human eyes, not machine calls. Tokens get burned. Mistakes get made. And the AI is still, fundamentally, only working with what it can see or infer, not with what is actually happening inside your systems right now.

The Third Option: WebMCP

Alex Nahas was a backend engineer at Amazon when the Model Context Protocol (MCP) standard landed in early 2025. Amazon, with thousands of internal services, spun up an enormous MCP server, and the result was chaos. Too many tools fighting for space in the context window. Authentication that nothing could agree on. A mess.

The real insight Alex had was simpler: run MCP inside the browser. At Amazon, all authentication already worked through a single federated browser session. By integrating MCP into a client-side JavaScript application and communicating via a Chrome extension, he built something that worked with the identity systems that already existed rather than fighting them. He called it MCP-B, the B stands for browser.

Teams from Google Chrome and Microsoft Edge were already thinking along similar lines. Google was prototyping something called "script tools." When MCP-B arrived, they joined forces through the W3C Web Machine Learning Community Group, and the result is now being standardised as WebMCP.

The idea is straightforward. Instead of an AI agent fumbling through screenshots or raw DOM, a website can explicitly expose what it can do, as JavaScript functions, with descriptions, with structured schemas, so that an AI agent can call them directly.

Imagine your website telling a visitor's ChatGPT or Claude agent: here is what I can do, here is how to ask me to do it.

That is the shift.

What WebMCP Means in Plain Terms

Without WebMCP, an AI interacting with your systems has to:

  • Take screenshots and guess what buttons do
  • Parse raw page code to find anything useful
  • Burn through tokens working out context that should have been handed to it directly

With WebMCP, the same AI can:

  • Call a clearly defined function: get the current device status, pull the route table, check service health
  • Navigate your application with a map it was given, not one it had to reverse-engineer
  • Submit changes, fill forms, trigger workflows, and pause for human approval before anything irreversible happens

The functions being exposed are not new. They are the same functions your web application already uses to submit forms, query data, and update state. WebMCP just makes them discoverable and callable by agents.

As Alex put it: even an old WordPress site, with the right tools registered, becomes "basically an AI company." No framework required. Vanilla JavaScript is enough.

WebMCP's Three Tool Types: Read, Navigate, Write

Through building WebMCP-enabled sites, Alex identified three distinct categories of tool that map cleanly onto how agents need to interact with any system:

Read-only tools are always available. These are your GET operations, fetching data, reading state, checking status. You surface everything and let the agent pull what it needs. No navigation required.

Navigation tools are the system prompt of your website. They tell the agent what the site does and where things live. Think of it as a blueprint: here are the links, here is where more tools will appear as you go deeper. These are marked as destructive because they cause visible changes (the page actually navigates), which means agents that are well-behaved will flag the action to the user first.

Write tools take action. They fill out forms, submit data, trigger changes. This is where human oversight matters most, and WebMCP supports this through an elicitation flow: the agent can prepare a change, surface it to the user with "here is what I am about to do, yes or no?", and only execute on confirmation.

For network operations teams, this maps directly onto the work. Monitoring is read-only. Navigating between your management tools and dashboards is navigation. Pushing a config change, opening a ticket, or triggering a remediation workflow, those are write tools, with appropriate human approval gates built in.

If you want to understand how this applies to your specific infrastructure, book a demo with us and we can walk through it against your environment.

WebMCP and Security: A Meaningful Step Forward

There is an honest security concern with AI agents operating on web systems that is worth naming directly. It has been called the "lethal trifecta": an agent with context from multiple browser tabs can potentially be manipulated by a malicious site to act on data from a trusted one. Tab A is your bank. Tab B is something hostile. An agent sitting in the middle treats both as equally valid sources of instruction.

WebMCP does not eliminate this problem entirely. But it meaningfully reduces the attack surface. Instead of the agent being able to see and act on everything a user can, screenshots, DOM, hidden elements, whatever, it can only call the tools the website has explicitly chosen to expose. You can hash tools to detect tampering. You can scope trust to specific domains with expiry times. You can build elicitation flows that require explicit human approval before anything consequential happens.

Going from "the model can see and do anything" to "the model can call these specific functions" is not a complete solution, but it is a significant narrowing of what is possible for an attacker to exploit.

WebMCP Timeline: Where Things Stand in 2026

WebMCP is early. The specification is moving from community incubation in the W3C Web Machine Learning Community Group toward a formal draft. Chrome has an experimental implementation available behind a flag in Chromium now. Official browser announcements from Google and Microsoft are expected around mid-to-late 2026.

Alex maintains a polyfill that developers can use today, install @mcp-b/global and register tools with navigator.modelContext.registerTool(), and it works with any frontend framework. There is also a fork of Chrome DevTools that lets you call WebMCP tools directly, which means an AI coding agent can write and test its own WebMCP integrations.

The direction is set. When this lands in browsers at scale, every website that has registered its capabilities becomes, in effect, an interface that AI agents can use as fluently as a human expert would.

What This Means for Network Operations

Most of the systems that network teams interact with daily are web-based. Monitoring dashboards. Device management portals. Cloud console interfaces. Ticketing platforms. Configuration management tools.

Right now, an AI assistant working with your team can read documentation, suggest configurations, and help reason through problems, but it cannot touch any of those systems. It is, in the end, still a text machine.

The automation infrastructure we help teams build is designed with exactly this trajectory in mind, clean APIs, structured workflows, tooling that an AI agent can actually interface with rather than just read about.

When WebMCP becomes standard, and the combination of Google, Microsoft, and W3C backing suggests it will, the organisations with that foundation already in place will be able to move immediately. Those without it will be starting from scratch.

If that conversation is worth having for your team, book a demo and we can show you exactly what AI-ready network operations looks like in practice. The systems you are building now will either be ready when this shift arrives, or they will not.

The shift from AI-as-reader to AI-as-operator is not five years away. Official browser support is expected before the end of 2026. The question is not whether to prepare, it is whether you are starting now or catching up later.

Further Reading

Leave a Comment

Need Expert Help with Network Automation?

Let's discuss how we can help transform your network infrastructure

Schedule a Consultation