Greg Isenberg just put out a video about a Google feature almost nobody is talking about, and I think he's early again.
It's called WebMCP. The pitch in one line: your website stops being a wall of HTML that AI agents have to squint at, and starts handing them a short menu of actions. Here's how to search. Here's how to book. Here's how to buy.
Why should you care? Because millions of AI agents are about to become a new kind of visitor on the internet, with money to spend, and almost every website they'll land on is built for human eyeballs only.
Greg brought his friend Vinnie on to demo the whole thing, and then pitched two businesses you could start on top of it this week. I took notes on all of it. Video first, my breakdown below.
What WebMCP actually is (in plain founder language)
Today, when an AI agent uses a website, it does it the ugly way. It downloads the entire DOM, scans through all the code, takes screenshots, and guesses where to click. It's slow, and it breaks the moment you move a button.
WebMCP flips that. The website itself exposes a small list of tools the agent can call directly. Vinnie's phrase for it stuck with me: it's basically websites with agent buttons.
If you know MCP (the protocol that lets AI models talk to external tools), WebMCP is that idea moved into the browser UI. It's an experimental feature from Google and Microsoft. The proposal went in two years ago, it launched in February, and right now you have to turn it on manually in Chrome flags.
Experimental, yes. But it's Google and Microsoft jointly deciding what agent-ready websites should look like. Greg's read: they built it because they don't want a broken internet where billions of dollars of agent-driven purchases fail at checkout.
That framing matters for where this sits in history:
→ SEO asked: can Google understand your page?
→ AEO asked: can AI cite your answer?
→ WebMCP asks: can an agent finish the job on your site?
That last question is the one that moves money.
The demo: an espresso shop an agent can actually use
Vinnie built the demo around a real annoyance from his own life: buying an espresso machine.
He was upgrading from his Mara X and needed a machine that could pull two flat whites a day and fit a 32 cm counter space. Anyone who's bought espresso gear knows the pain. Specs everywhere, accessories in incompatible sizes, endless tabs of "will this portafilter fit?"
So he built a demo espresso store with WebMCP tools baked in. Sixteen of them.
He typed his requirements to his agent, and the agent didn't scrape anything. It called the store's tools, pulled two matching machines, laid them out side by side on the actual page, and highlighted exactly which specs matched his request. Width, capacity, portafilter fit.
Then it got better. He asked which accessories fit the Bianca the store had suggested. The agent already knew his order history through the store's get-my-gear tool, recommended a water softener two-pack, added it to the cart when he said yes, and applied a 10% coupon at checkout on its own.
And here's the part founders should slow down on: there are no API keys and no MCP server to configure, because the tools ride the browser session. Logged out, the demo store shows 3 tools. Logged in, the full list unlocks. Your customer just logs in like a human, and their agent inherits the access.
Compare that to every other way agents touch your product right now. Raw APIs and MCP servers are headless and make users juggle tokens. Computer use is slow and fragile. Browser automation chokes on your HTML. And the in-app chatbot you were about to build? There's a sentiment flying around on X that Greg pulled up, paraphrasing a Cloudflare exec: people don't want to use your product's agent, they want their own agent to use your product.
WebMCP is the middle path. The user keeps their agent (and its memory and context), you keep your UI, and the two finally speak the same language.
Where this hits first
Vinnie called out four zones where WebMCP makes immediate sense:
→ Compatibility-driven commerce. Espresso gear, camera systems, car parts, dev boards. Anywhere "does this fit with that?" kills conversions.
→ SaaS admin panels. Analytics, marketing tools, settings consoles. Let your customer's agent drive the dashboard instead of building your own sidebar bot.
→ Regulated industries. Banks and insurers can expose read-only tools that guide users to the right place without ever handing an agent write access or an API key.
→ Internal tools. Before you touch anything customer-facing, make your own back office agent-ready for your team.
I'd add a fifth: anyone selling to founders. Your buyers are the most agent-native people on earth. They will absolutely send an agent to compare you against competitors. The question is whether your site helps that agent or fights it. I wrote about this exact shift in selling to AI agents, and WebMCP is the first real infrastructure answer I've seen.
Business idea #1: the WebMCP conversion agency
Greg's first startup idea is a classic picks-and-shovels play: an agency that makes boring business websites agent-ready.
Law firms, HVAC companies, med spas, dentists.
None of them will ever read a Chrome spec. But they already see agents showing up in their offices, and they'll understand the pitch in one sentence: a new class of visitor is coming to your website, it books jobs and requests quotes, and right now your site is invisible to it.
The money model Greg sketched:
→ Setup: $200 to $10,000 depending on the client, building their V1 tools (request a quote, book a consult).
→ Retainer: a few hundred up to $700 a month for monitoring and upkeep.
That retainer is the business. Recurring cash flow from clients who will never rip it out, in a skill almost nobody has yet. Vinnie's tweak: don't sell "WebMCP" at all. Sell an agent conversion agency, and let WebMCP be one tool in the package alongside MCP servers and whatever the client's setup needs.
This rhymes with the agent-economy businesses I broke down in 3 agent internet businesses to start now. The infrastructure keeps changing. The "translate new tech for boring businesses" playbook never does.
Business idea #2: the agent mystery shopper
Idea two is even simpler, and honestly my favorite.
You send AI agents to complete the most important user journeys on a client's site. Buy a hoodie from the Shopify store. Book the consult. File the claim. Then you hand the client a report: here's where the agent got stuck, here's the bad product description that confused it, here's the missing tool that cost you the sale.
Greg compared it to a credit report, but for agent readiness. Charge $100 to a few hundred per month, run it weekly or monthly, and every repeated fix you spot becomes a feature of a future software product.
You could stack it on top of the agency, too. The mystery shopper report is the audit that sells the conversion work. One finds the leaks, the other plugs them, and both bill monthly.
"But it's experimental" is exactly the point
The obvious pushback: WebMCP is an experimental Chrome flag. Why build anything on it?
Greg's answer is the one I'd give you about any early channel: the arbitrage exists while it's experimental. When this stuff ships by default and every agency offers it, the window is closed.
Vinnie's version was the GPT analogy. People played with GPT-2 and GPT-3 and shrugged, "years away from being serious." That was roughly two years before ChatGPT detonated. He's not fading a standard that Google proposed two years ago and has been shipping steadily since.
Both of them expect that within 24 to 36 months, basically every consumer will have their own agent. If they're even half right, "can an agent use your website?" becomes a line item on every marketing audit.
You don't need to bet the company. Enable the flag, clone Vinnie's demo repo, and spend one Saturday making one journey on your site agent-ready. That's the whole cost of being early.
FAQ
What is WebMCP?
WebMCP is an experimental browser feature from Google and Microsoft that lets a website expose a list of tools (search, book, add to cart, apply coupon) that AI agents can call directly. Instead of scraping your page's code and guessing where to click, an agent gets a clean menu of actions and can complete tasks on the live page while the user watches.
How is WebMCP different from an MCP server?
An MCP server is headless: it runs alongside your backend, and users have to connect it and manage authentication. WebMCP lives in the browser page itself and uses the existing session. If your customer is logged in, their agent has the same access, with no API keys or tokens to pass around. It also lets the agent manipulate the visible UI, so the human gets visual confirmation of everything.
How do I try WebMCP right now?
You need Chrome. Go into Chrome flags and enable WebMCP support, then allow remote debugging in Chrome inspect. Vinnie's demo espresso store (with its GitHub repo linked on the page) is open for anyone to clone and experiment with.
Is it too early to build a business on WebMCP?
It's early, and that's the opportunity. The spec launched in February and is backed by Google and Microsoft. The two models Greg pitched (an agent conversion agency and an agent mystery shopper service) both charge for outcomes businesses already understand, so even if the standard evolves, the service layer keeps billing.
Steal ideas like this every week
I pull apart videos and interviews like this one every week, plus I interview bootstrapped founders doing $100K to $10M a year about what's actually working. If you want the playbooks before they're obvious, the podcast is where I put them.
