There's been this ongoing challenge with AI agents: when they visit a website, they're basically tourists who don't speak the language. Whether you're using LangChain, Claude Code, or tools like OpenClaw, your agent is stuck guessing which buttons to press, scraping HTML, or processing thousands of tokens worth of screenshots just to figure out what's on a page. If you've been building with agents for a while, you know exactly how painful this is.
That's what makes Google Chrome's new WebMCP preview so interesting. Earlier this week, the Chrome team shipped an early version of what could be the most important change to how agents interact with the web in years. Instead of treating every website like a foreign language that needs translation, WebMCP lets websites expose structured tools directly to AI agents. No more scraping. No more processing endless screenshots. Your agent just calls functions.
This is part of a bigger shift we're seeing where the web itself is becoming more agent-friendly, not just more human-friendly. And honestly, it's about time.
I've been following the development of browser-based agents, and WebMCP caught my attention because it solves problems most people aren't even talking about yet. Watch my YouTube video on it below.
Why Current Web Interaction Is So Inefficient
Right now, agents interact with websites in two main ways. The first is through screenshots—you take an image of the page, feed it to a multimodal model, and hope it can identify buttons, form fields, and interactive elements. The problem? You're burning through thousands of tokens for every single image you process.
The second approach is accessing the DOM directly and parsing raw HTML and JavaScript code. While this uses fewer tokens than images, you're still translating from one language to another. The agent has to sift through paragraph tags, CSS styling, and all sorts of presentation markup that doesn't actually matter for understanding what actions it can take.
Both methods feel like working through a translator when you could just speak the same language.
How WebMCP Actually Works
The idea behind WebMCP is beautifully simple: let each webpage act like an MCP server that agents can query directly. The page basically tells the agent, "Here's what you can read. Here's what you can click. Here's what you can fill in."
This isn't entirely new—academics and companies have been proposing versions of this for a while. But in the second half of last year, Microsoft and Google actually got together to build a real spec for how this would work. The timing makes sense too—this was right around when we saw Perplexity release Comet and OpenAI release Atlas, when web interaction was clearly heating up.
What makes Chrome's approach interesting is that it's designed for human-in-the-loop workflows first. The agent works with the user, not just autonomously. So normal people still use websites normally, but agents can help speed things up and improve the experience.
Google presented three core pillars at the Web AI Summit: context (understanding what the user is doing beyond just the current screen), capabilities (taking actions on the user's behalf), and coordination (managing the handoff between agent and user when needed).
The Two APIs You Need to Know
Chrome has structured WebMCP around two main APIs. The Declarative API handles standard actions—think HTML forms with added tool names and descriptions. If you've already got well-structured forms on your site, you're apparently about 80% of the way there.
The Imperative API is for more complex, dynamic interactions that require JavaScript execution. This is where you'd define custom tools, similar to how you'd structure function calls for OpenAI or Anthropic's API endpoints.
The practical difference here is huge. Instead of dozens of interactions clicking through filters and scrolling pages, a single tool call could return structured results. Imagine your agent calling a "search products" function and getting back organized data instead of trying to parse a visual search interface.
What This Means Going Forward
While WebMCP is still behind a flag in Chrome, it's already in the browser. This isn't a theoretical spec anymore—it's actually happening. Google will likely roll this out fully at Google Cloud Next or Google IO in the coming months, and I expect things to move quickly from there.
We'll probably see tools and maybe even Claude skills that help convert existing websites to expose their own WebMCPs. For anyone building AI agents or websites that want agents to use them, this is definitely something to have on your radar.
The shift from agents guessing their way through the web to websites speaking the agent's language directly? That's the kind of change that makes everything else possible.
No comments:
Post a Comment