Picture this: An AI agent walks into a website. Well, not walks, more like stumbles blindly while yelling "WHERE'S THE SUBMIT BUTTON?"
The web was built for humans. AI agents are just...improvising.
The Current State: Digital Archaeology
Think about how you use the internet: You open a browser, search for something, click around a few pages, fill out some forms, maybe compare prices across tabs.
Perfect workflow for a human brain. Absolute nightmare for an AI.
When an agent tries to do something simple, like finding flights, comparing products, or booking a table, it has to either scrape HTML or literally analyze screenshots to figure out what's clickable.
It's the digital equivalent of trying to use a microwave by staring at it really hard and hoping something happens.
Agents don't "see" a booking form. They see a wall of div soup and CSS classes named things like btn-primary-v2-final-ACTUAL. Then they guess. And sometimes they guess wrong.
Enter WebMCP: The Web, But With an API Brain
This is why Web Model Context Protocol (WebMCP) caught my attention.
Instead of forcing AI to reverse-engineer your entire DOM structure like some kind of digital archaeologist, WebMCP lets websites expose structured actions directly to agents.
Not "find the button somewhere in this React component hell."
More like:
searchProducts(query, filters)bookFlight(origin, destination, date)cancelSubscription(subscriptionId)
Clean. Explicit. No hallucinations about whether the checkout button is actually a link styled as a button.
What This Actually Means
Websites stop behaving like static pages you have to scrape. They start acting like services that AI can talk to directly, the same way APIs work for developers today.
The web slowly evolves into something new: an environment designed not only for humans browsing with their eyes, but also for AI agents working on our behalf in the background.
We're not replacing the visual web. We're adding a second interface layer, one that speaks machine.
The Part Nobody's Talking About
Here's the thing: this only works if websites actually adopt it.
And we all know how long it took to get everyone on HTTPS. Or responsive design. Or accessibility standards.
So yeah, WebMCP is a clever idea. Whether it becomes the standard or just another protocol that 47 sites implement while the rest of the internet keeps doing <div class="button"> is the real question.
Check it out: https://webmcp.link/
Have you built anything with AI agents trying to interact with the web? What broke first, the scraper or your patience?
Top comments (2)
This is such a perfect framing of the problem. The "digital archaeologist" analogy is spot on — most AI agents are basically doing the equivalent of reading a restaurant menu by analyzing the pixels of the sign outside.
What's interesting is this same pattern shows up beyond web browsing. Email is another space where AI tools are forced to work against the grain. Most email AI products try to replace your entire email client (become the new UI layer), when what they really should be doing is working with the existing APIs.
Gmail has a well-structured API. So instead of building yet another email client that forces users to switch, the smarter approach is to use the API to read emails, generate draft responses using AI, and let the user review before sending — all through the existing interface.
That's actually what I've been building. A Mac app called Drafted that uses Gmail's API + your own AI key (OpenAI, Gemini, or Claude) to pre-draft replies. It never auto-sends anything. Runs 100% locally. The user is always in the loop.
Your point about WebMCP exposing structured actions is exactly the right direction. The best AI tools don't fight the existing infrastructure — they work with it.
What domains beyond web and email do you think are most ripe for this "API brain" approach?
Some comments may only be visible to logged-in visitors. Sign in to view all comments.