I recently wrapped up a new course for LinkedIn Learning, WebMCP in 10 minutes, just in time for a blitz campaign by OpenAI and partners (including Google Chrome, Cloudflare, Shopify, Netlify, Vercel, Render). (Unfortunately, making the course meant I didn’t have time to enter the competition, but I’m super excited to see the winners on Sep 23rd!)
Why the sudden hype? Although it's available in a Chrome origin trial through Chrome 156 and in ChatGPT's built-in browser, WebMCP is just a Draft Community Group Report from the W3C Web Machine Learning Community Group, which in W3C process terms means it is not a W3C Standard, nor is it on the W3C Standards Track. The group only resolved in May to begin early wide review with the TAG and the Privacy and Security groups. The draft is also still moving under implementers’ feet. provideContext() and clearContext() were cut in March, standalone unregisterTool() gave way to AbortSignal weeks later, and the whole API moved off navigator onto document. Microsoft co-authored the spec, then buried it behind a flag in Edge, which raises questions about how seriously they're investing in an agentic web or how confident they feel in the spec.
If you’d like to learn more about where WebMCP comes from, I interviewed one of the people behind it, Alex Nahas.
The user’s right to their browser session
You may remember that OpenAI launched Atlas, an Agentic browser, then shortly after retired it. But Atlas was technically folded into ChatGPT itself, as a sort of browser mitochondria to let the agent escape hatch into the web to get things done on behalf of its users, like accessing LinkedIn contacts or placing orders on Amazon.
Which raises the question of whether you’re allowed to. Amazon sued Perplexity in November 2025 over people shopping through Comet, Perplexity’s agentic browser, and won an injunction blocking the browser from its site only to lose it on appeal on August 4, 2026. The case isn’t over, and the future of what we can and cannot do with AI Assistants and agentic browsers on the web hangs in the balance.
Don’t mistake Amazon’s lawsuit as a stand for users’ experience and wellbeing. Today’s retailers want to maintain control over as much of the purchasing surface area as they can so they can upsell the user and sell ads to distract them. Giving that up to an agent is a threat to a lucrative business model (even if it provides a better experience for the customer). The Ninth Circuit agrees that when a user directs the agent, the access is the user's, not Perplexity's. That’s important for the future of the web.
WebMCP as a bridge to the Agentic Web
So it’s no surprise that WebMCP would be celebrated with the full gusto of the two companies who have the most to gain by making the web easy for agents to use. Google backs a future where people use an agentic Chrome and/or Gemini to traverse the Web that it helped cultivate for so many years, and OpenAI is betting on ChatGPT being the window to the world people use. In both cases, the web becomes a sort of human-first fallback layer while people migrate to agents.
This is not a bad thing for the consumer. Imagine if an agent could fill out your tax forms! Goodbye, TurboTax! There will be winners and losers as agents rise in usage. WebMCP requires site owners to opt into making their content easier for agents to access. Shopify is doing just this, taking the opposite stance from Amazon and actively embracing agentic shopping. Shopify has been actively exploring MCP Apps and WebMCP as part of agentic shopping cart experiences with OpenAI. And this makes sense for Shopify, as their customers are merchants, not shoppers. When people just want to buy things, the fastest route to the best price and highest quality historically wins. This naturally worries direct providers like Amazon as it levels the playing field by standardizing the user experience across one agentic surface.
WebMCP doesn’t increase your site’s discovery, but it smooths out its operability for agents by offering them a selection of on-page tools. Tools are tab-bound and ephemeral, so an agent has to already be on your page to know they exist. We don’t need WebMCP for agents to access and manipulate web sites. But WebMCP makes it faster and easier for agents to do things without sending multiple screenshots and DOM dumps to data centers. WebMCP cuts down roundtrips by making each page on a site able to surface a set of tools to visiting agents so long as they’re on the page.
These tools can be the same JS functions that you’d normally trigger via human clicks and interactions, or they can be tied to forms humans fill out. Essentially, you’re exposing a function, a description, and a schema shape and the agent calls the tool without having to make a dozen inferences and invoke a bunch of synthetic events. It’s more efficient and more accessible for agents. (If you want agents to be able to do things on your site, that is.)
Security sidebar
On the security side of things, you might be concerned about tools running inside the page, with access to your user's live session and cookies. An agent that has read something untrusted can be talked into calling them. Prompt injection is not a solved problem. What you get today is two annotations: readOnlyHint, to flag tools that don't change state so the agent knows when it should be asking permission, and untrustedContentHint, to label output that came from somewhere you don't control (comments, reader submissions, anything syndicated). requestUserInteraction() is a method for a tool to stop and ask the human mid-execution, but is still in the IDL with no normative algorithm behind it. Consent management is still an open discussion in the working group.
Should you make your site agent-ready?
Will WebMCP change how the web works? I suspect it’s a bridge technology that we’ll see gain adoption as users adopt AI assistants. Using this technology is going in on this bet with Google and OpenAI.
Before you adopt, consider if your users are starting to use AI assistants like ChatGPT or agentic browsers like Comet. Do you want it to be easier for them to get things done on your site? If yes to either, trial WebMCP on one important flow and watch the usage over the next three to six months.
Fortunately, if you already have semantic, accessible forms, making your site Agent Ready is as easy as adding a few attributes. It is easy to hedge your bet and accommodate the growing ranks of agent-first users.
