Rethinking the Browser

Over the past year, a new category has quietly emerged: the AI browser. Unfortunately most of them are just browsers with an AI assistant bolted on.

So I asked myself: What would a browser look like if it were designed from the ground up for AI? The line between search engine, AI assistant, and web browser should be completely removed.

What an AI Browser Should Be

1. Privacy First

AI shouldn't require surveillance. The browser runs a lightweight LLM locally, so the user's data never leaves the machine. Search is anonymous or even based on a local index. And the AI automatically rejects non-essential cookies because privacy shouldn't depend on how many cookie banners we're willing to click.

2. Browser as Search

Why have separate steps for searching and opening results? When a user searches, the browser immediately opens the top results as tabs, giving the user real content instead of a blue-link list. The user is always zero clicks away from the results.

Most websites have narrow layouts limited to 800 pixels. That is much less than the typical monitor, and wasting a lot of screen space. We could have pages living side-by-side on a horizontally scrollable container, not on top of each other. That way, the users can compare information more easily.

Do we still need a search engine? We could simply use a local index for 99% of queries. A 128GB index could store tens of millions of pages. Users could just download their favourite index from somewhere or even mix and match.

3. Concise AI Assistant

A bottom bar shows only the latest AI response, short and to the point. It's not a wall of text, it's a quiet, contextual companion. When users need more, they can ask for it, expand it, look at the chat history. Otherwise, it stays out of the way.

4. Observant but Private

The AI observes what users are doing (privately, with their consent). It remembers their open tabs, the context of their research, and how they're interacting with the pages. It understands context: what users are trying to do and how to help them achieve it. It learns from how users use websites and remembers their preferences over time. It alsow allows them to view the information it gathered and edit or delete it if they want.

5. Agentic Behavior

This isn't just a chatbot. When it notices patterns, drafting an email, researching a topic, comparing docs, it could offer help before users even ask. The assistant can interact with web pages on their behalf: fill out forms, extract content, take notes, trigger actions or even run commands in their terminal, similar to Claude Code.

Try the Demo

Here is a mock UI of the browser. You will need to open it on a desktop monitor.

Extensibility

But that's not all. This browser is designed to be extended in various ways.

Content-Type Plugins

Any page can be:

or anything else a plugin can provide.

The browser becomes a unified workspace, not just a shell for HTML.

Search Plugins

Search isn't limited to the web. Through search plugins, we can query:

...or any data source we can connect.

The AI decides which search providers to query, how to prioritize results and remove duplicates.

MCP Servers

Through MCP (Model Context Protocol) servers, the AI can gain new powers:

MCP turns the assistant from an observer into an active participant in the digital environment of the user.

AI-Friendly Websites

We could define a standard so that websites could embed simple meta elements to provide structured instructions for AI agents:

This opens the door to a new class of AI-aware web experiences that work with the agent, not against it.

Why This Matters

Browsers have barely changed in 20 years. Search happens somewhere else. AI happens in another tab. Our tools live in silos.

This project imagines a future where your browser is:

And it does all of that while respecting the user's privacy.