I’ve been thinking about how AI desktop clients like ChatGPT and Claude are expanding beyond chat. With MCP, MCP Apps, and now WebMCP, they could start interacting more directly with services and websites.

I gathered a few thoughts about what that might mean for browsers and wrote a short post about it.

Especially LinkedIn + AI Slop Posts are just unbearable.

Looks Cool! Navigation is good overall. But It would be great if I could use ArrowUp/ArrowDown to scroll in scrollable areas and focus on buttons as well.

Nice Project! You can also keep track of the recordings that the user already listened and pick the random from the ones user hasn't listened yet.

Thanks! That’s a great addition. I’ll work on adding that. Appreciate the feedback.

A few hours ago. I am learning a new language (rust) so I disabled auto complete and AI suggestions. Otherwise AI becomes more of a distraction than a booster.

But for languages I am more confident (like JS and dart), I usually don't write code and only review it.

If it is a throw-away project (like I am only experimenting something or just need something for one time), I don't event look at the code and only focus on the outcomes.

Interesting, I just decided (well 2 days ago) that I want to learn Rust myself. What resource are you using to learn?

I am using Rust By Example Book. https://doc.rust-lang.org/stable/rust-by-example/

Thanks, I appreciate your sharing this info.

I've seen a YC startup working on this. https://silahq.com/

As the article states, it is a "wild experiment". I wouldn't let AI control anything serious end to end. Also if Claude really becomes JIT, it is going to be an expensive one.

The idea is interesting though.

LoL :) that makes sense, but what if there is a new AI model releases when I am offline.

You will not suffer much from not adopting a new AI model for two weeks after release.

Hi HN! I'm the dev behind Agent Browser.

I was building AI agents that browse the web and kept hitting the same wall: sending screenshots or accessibility snapshots to the model burns tokens and is slow. So I tried rendering pages as ASCII wireframes instead.

Every interactive element gets a numbered label. The agent sees a compact text representation and says "click [12]" or "type [5] hello". A typical page costs 3x to 10x less tokens depending on the content.

You can use it with the Vercel AI SDK or as an MCP server. To add it to Cursor or Claude Desktop:

{

  "mcpServers": {

    "agent-browser": {

      "command": "npx",
      "args": ["-y", "@agent-browser-io/browser", "mcp"]
    }

  }

}

Still experimental, would love your feedback. Happy to answer questions!

Github stars started feeling more and more meaningless every day.