Andy Halliday

You could say, filter all of the search results on this page.

Or you could create a web ticket on GitHub or any of those things.

Instead of having to guess which DOM elements you have to act on and then click in the correct order, no.

Now the website has its own way of saying, here are the high-level things that you can do on this website, and here are the commands to invoke each one of those.

How do you put it out that efficiency advantage?

Because WebMCP from Google can perform at a very high accuracy level.

So like a checkout could be completed with 98% accuracy where a browser agent might flub it halfway through.

But by shifting from examination of pixels to this structured schema approach,

WebMCP cuts the computational overhead by 70%.

Yeah.

So it uses only 30% of the tokens, basically, to do that.

You can imagine how token-intensive it is for a web browser to be using AI to try to figure out what the pixels on that page are.

Yeah.

And isn't it...

So let me describe specifically how Cloud Code looks at my other tabs through the Cloud MCP view.

And it says, okay, wait, let me take a snapshot and take a look at lovable.

So I'm working in lovable over here.

So it's looking at lovable.

And then it says, okay, I'm going to scroll down in the chat and see what else was discussed.

Well, I guess was and then I can see it jump around like the browser is now moving around because Claude over here is working it in order to understand what the conversation prior to where I kind of ended up on the page was.