evolve(browser)

Agent Flow Diagram
GIF
Example of one of our users easily creating an extension to sort through their spam email.

Inspiration

Web browsing is static. Browsers and websites treat all of their users almost exactly the same, but there isn't a one-size fits all for the billions who use the internet. If a user wants some feature added or changed (e.g. hiding all YouTube shorts when visiting the website to avoid distraction), they must either: (i) hope the browser/website developer can and are willing to make it for them, (ii) hope and trust someone made a Chrome extension for their exact use case (without charging an arm and a leg), or (iii) have the time and expertise to make it themself.

We believe we need to stop constraining users with this one-size fits all mindset. As such, we built evolve(browser), a fork of Chromium (and Chrome Extension) which makes customizing your browsing experience as easy as sending a single message.

What it does

evolve(browser) pairs an advanced coding agent with complete browser context. This AI agent codes extensions which are loaded directly into the user's browser in real-time. Here are some cool extensions we created for ourselves or that our friends requested using evolve(browser) (these all took only 1-2 prompts):

Alastair's YouTube filter: When visiting YouTube.com, hide all YouTube shorts and all videos which have any League of Legends keywords in their title between 9 AM - 5 PM.

youtube_extension_gif

Haley's Email Manager: Stanford students get a lot of spam (and are forced to use Outlook), Haley wanted an extension which will help her quickly delete all her spam emails.

email_manager_gif

Antonio's Ad Blocker: Chrome banned uBlock Origin (Antonio's favorite ad blocker), but Antonio doesn't want to switch browsers. Antonio uses evolve to create a custom ad blocker for his most visited sites (and if he ever wants to extend its ad-blocking capabilities to new sites he can just tell evolve to update the extension).

adblocker extension gif

How we built it

Agent Flow Diagram

Very high-level agent breakdown:

A primary large model (GPT-5.2 or NVIDIA Nemotron Super 49B) performs the main reasoning, deciding whether to answer directly or call tools.
Context tools enhance responses using graph-based RAG over the extension codebase and, when useful, live browser context retrieved via WebSocket.
The knowledge graph is built offline by chunking the codebase, extracting entities and relationships with GPT-5-Nano, and generating embeddings with NVIDIA’s embedding model.
When actions are required, coding tools (sandboxed terminal, linter, testing sandbox, and code editor) execute tasks, sometimes assisted by a smaller secondary LLM.
The primary model synthesizes everything into the final response, which is streamed back to the chat and auto-reloaded browser extension.
A secondary model extracts rules / memories from the overall conversation for future conversations about this extension.

Challenges we ran into

Our fork of Chromium has limited functionality currently because build times take so long (we weren't able to iterate quickly)
For the Chrome Extension, auto-loading / auto-reloading chrome extensions from extension code is guarded against by Chrome due to security considerations, so we created a hacky solution (utilizing Apple Events to insert JavaScript directly into Chrome).