What Web MCP means for your website

On 10 February 2026, Google and Microsoft submitted a draft standard to the W3C called Web MCP (Web Model Context Protocol). It changes how AI agents interact with websites, and it matters to anyone whose business relies on people finding and using their site.

This is not yet in production. Chrome 146 Canary supports it behind a flag. The spec is incomplete in places. But the direction is clear enough that it is worth understanding now, before it becomes infrastructure.

How AI agents currently use websites

When an AI agent (Claude, ChatGPT, Gemini) needs to interact with a website today, it has two options. It takes a screenshot and runs it through a vision model to figure out what to click, or it scrapes the DOM and guesses at structure. Both approaches are slow, expensive and unreliable.

Screenshot-based agents typically achieve around 70% task accuracy. Every interaction burns thousands of tokens. Authentication is handled through workarounds: separate API keys, OAuth flows layered on top. There is no standard protocol, so every agent framework solves the same problem from scratch.

This is the baseline your site is operating against right now when AI agents try to use it.

What Web MCP changes

Web MCP introduces a browser-native JavaScript API, navigator.modelContext, that lets any website register structured tools directly in the browser. An AI agent discovers those tools, then calls them with structured JSON input. No screenshots. No DOM guessing. The browser sits in the middle as a trust layer, enforcing permissions and managing consent.

The performance difference is significant. Compared to screenshot-based approaches: 89% improvement in token efficiency, 67% reduction in computational overhead, task accuracy rising from around 70% to 98%.

That is not a marginal improvement. It is a structural change in how agents interact with the web.

Your existing forms can become agent-callable tools

The simplest implementation requires no JavaScript. Web MCP offers a declarative approach: add a handful of HTML attributes to an existing form (tool name, description, whether to autosubmit) and the browser handles the rest. It reads your input fields, generates a JSON schema automatically, and the form becomes callable by any AI agent that supports the standard.

For more control, the imperative API lets you register tools via JavaScript. You define a name, description, input schema and an execute callback. That callback runs in your page's JavaScript context, which means it already has access to the user's session and cookies. No separate authentication layer.

A travel site could expose a flight search function. An e-commerce store could expose product filtering and checkout. A support tool could expose ticket submission. These are things your site already does. Web MCP gives AI agents a clean, structured way to call them.

The business case is straightforward

Customers are already using AI tools to find, research and buy things online. I have written previously about how customers find products with AI and what it means for your store. That shift is happening regardless of what you do to your site. Web MCP determines whether AI agents can interact with your site reliably when customers ask them to.

A site that supports Web MCP gives agents structured, accurate tools to work with. A site that does not gives agents a screenshot to guess from. The gap in reliability is the gap between your site completing a task and failing it.

For e-commerce this is concrete: an agent that can call a structured search and checkout tool will complete purchases. An agent forced to scrape and click will drop off. For lead generation, an agent that can call a structured enquiry form will submit it. An agent trying to navigate a CSRF-protected form without session context will not.

How this relates to Anthropic's MCP

These are different protocols that solve different problems. Anthropic's MCP is a backend protocol (JSON-RPC over standard I/O) connecting AI platforms to server-side tools and data sources. Web MCP is a frontend protocol, a browser-native API connecting agents to client-side web pages.

Together they cover the full stack: backend and frontend, server and browser. A business running both would have structured AI access to their internal systems via MCP and structured AI access to their public website via Web MCP. Neither replaces the other.

What is not ready yet

The spec has gaps. Security sections covering prompt injection through tool descriptions and data exfiltration through tool chaining are still marked as placeholder. These are real concerns and they need real answers before this becomes a standard you would deploy on a production site handling sensitive transactions.

Browser support is also limited. Chrome 146 Canary is the only implementation. Edge support is plausible given Microsoft's co-authorship, but Firefox and Safari have not indicated plans. Broad adoption depends on that changing.

The standard was drafted in February 2026 and is still a W3C Community Group draft, not a formal W3C recommendation. The path from draft to recommendation can take years.

What you should do now

Nothing urgent, but a few things worth tracking. First, understand what your site's current AI accessibility looks like: which forms and functions exist, and how hard they are to interact with via automation. Second, watch for Chrome stable and Edge shipping navigator.modelContext; that is when implementation conversations become practical. Third, consider whether your site's current structure would make Web MCP implementation straightforward or difficult. Clean, semantic HTML will make the declarative approach much simpler.

The best AI integrations are ones that work reliably and do not require constant maintenance. I have written about that principle before and it applies here too. Web MCP is a structural improvement to how agents interact with websites, not a feature you bolt on, but a standard your site either supports or does not.

If you want to understand what this means for your specific site, what your current AI accessibility looks like and what it would take to be ready, that is something I can assess as part of an AI integration review.