Home / Blog

Broxee meets the real web

We started using Broxee for everyday browsing. Here is what broke, what we fixed in the engine, and what still needs work.

Pandhu Wibowo November 17, 2026 5 min read

Until this week, most of what we knew about Broxee came from its test pages and a budget script. Then we did the obvious thing and used it: searching, clicking results, opening links in new tabs, switching tabs while pages were still loading. Real sites found problems the test pages never did.

This is the whole window, chrome included. The sidebar, the header and the page are all drawn by Broxee's own engine.

Small things a browser is expected to do

Some of the work was just catching up with habits everyone has:

Tabs that keep loading in the background

Broxee keeps only the visible tab in memory as a live page. Everything else sleeps in a file on disk, which is how it stays under 60 MB. Until now, switching away from a tab that was still loading simply stopped the load.

Now a tab you leave keeps downloading: the HTML, stylesheets, scripts and images, exactly as they arrive. What it doesn't do is build the page, because a built page costs 10 to 20 MB and two of them at once could break the budget. When the download finishes, the tab goes to sleep with its file, and its title in the sidebar updates. When you come back, the page builds from that file without asking the network again.

A page that went blank below the first screen

One payment site rendered its first screen and then nothing at all when you scrolled. The cause was a single CSS rule that many sites built with Tailwind and Next.js have: the root element is as tall as the window, and it clips overflow sideways.

Broxee was clipping to the root element's box in both directions, so everything below the first screen was cut off. Browsers treat the root's overflow as belonging to the window instead. Broxee does now, and a test keeps it that way.

Images that scripts add

Search results showed empty circles where site icons should be. The icons weren't in the HTML at all: the search engine's script adds them after the page loads. Broxee only ever loaded images that were in the HTML when it arrived, which also broke the common "lazy loading" pattern where a script fills in an image's address later.

Now every image is kept by its address, and an image shows the file of its current address, whoever set it. Images that arrive after the page has loaded are fetched in the background, and the page is laid out again when they come in.

JavaScript that sites actually use

Broxee's JavaScript interpreter is written in the repository, and real pages kept stopping on things it didn't have yet:

How we test a browser we can't click

Broxee's tests never send keystrokes or clicks to the real desktop, since someone is usually working on it. Instead, the browser runs against an in-memory screen, and a test renders the window to images: the sidebar, menus, dialogs, split view, light and dark. That's how we found and fixed one visual bug this week, the pane number box missing from the active tab in split view. It also produced the screenshot above.

There is also a regression suite of 22 saved pages that must render exactly the same bytes on every run. Three of them changed this week, all on purpose: a skip link that used to cover the Go logo is now hidden, as in other browsers; SQLite's site shows its search button now that location.origin exists; and the Rust documentation no longer applies a stylesheet meant only for browsers without JavaScript.

What's still rough

The binary is still 3.7 MB, and the stress page still peaks at about 50 MB, under the 60 MB limit.

Everything here is in the Broxee repository, with a test for each fix. If a site looks wrong in Broxee, open an issue with the address. Those reports are exactly how this week's list was made.

All posts