We built ego, a browser where humans and AI agents work side by side

David — Mon, 27 Jul 2026 14:37:25 +0000

TL;DR

We built ego lite, a Chromium-based browser where you and an AI agent share the same browsing environment. You use the browser normally while agents run web tasks in parallel inside their own isolated spaces, reusing the sessions you're already logged into. It's free and macOS-only for now.

Site: lite.ego.app

GitHub: citrolabs/ego-lite

This post is partly a launch announcement. But mostly, it's us thinking out loud about the problem and what we ended up building.

Why build a browser in 2026?

The short version: this came out of our own daily grind.

Over the past year, most of our team's time has gone into working with agents like Claude Code and Codex. They've moved well past writing code—they do research, wire up SaaS tools, and get real work done. But a lot of that work only exists on the web. Plenty of services don't expose an API. The actual information and functionality just live in the GUI, full stop. So if an agent wants to finish the job, it has to drive a browser itself.

And browsers just weren't built for that. Point an agent at one and you hit the same two walls every time: it can't access a session you're already logged into, so it stalls on login screens and CAPTCHAs before doing anything useful; and once it is running, it takes over your window and mouse, so you can't use the machine for anything else while it works.

What we actually wanted was simple: a browser you can use every day, that an agent can plug straight into, and that stays out of your way while it works.

That's what ego lite is.

It's a Chromium fork. On first launch, it imports your bookmarks, extensions, cookies, and login state from Chrome, so it slots in as your daily driver with zero friction. During installation, it also drops an ego-browser skill into your agent's skills directory—no extra setup is needed to start driving it.

When an agent goes to work, it runs in its own Space: a separate workspace that reuses your logged-in state without touching the tabs you're actually using. You can watch it the whole time and take over whenever you want.

Here's a quick demo if you want a sense of how it works: Watch the demo on YouTube

One more thing worth mentioning: model pricing has been in freefall lately, with Anthropic and OpenAI both cutting token costs as they compete for market share. A \$200-a-month plan is basically impossible to burn through on coding alone anymore. So pointing the leftover budget at browser tasks just makes sense.

Here are the two workflows we lean on most ourselves

Automating long-running tasks

Long tasks fail in two boring, predictable ways: they get stuck partway through, and they hog the browser the entire time they're running.

Spaces are built to solve both problems. A Space isn't a new window, a Chrome profile, or headless mode—it's an isolated workspace inside the browser itself. The agent works in its own Space, you work in yours, and neither one steals focus from the other. If the agent hits a login wall or a CAPTCHA, you just switch to that Space, clear it, and let the agent pick up where it left off.

Because the agent writes out the whole script up front and runs it in one shot—instead of poking at the page step by step—long workflows are more stable and use fewer tokens. The idea is to collapse a multi-step task into a single execution instead of looping through "run two commands, check the result, run two more." Compared with traditional CLI-driven approaches, it's 20–50% faster on complex workflows, with higher success rates and far fewer tool calls per task. We saw the same pattern in our internal benchmarks against Vercel's agent-browser: the more complex the task, the bigger the gap.

It also supports parallel execution, which is honestly the part we use most. You can have Claude Code sorting leads across 10 Spaces while Codex scrapes competitor pages in five more, and none of them interfere with one another. In head-to-head tests against Vercel's agent-browser across four complex tasks, our best result was 2.6× faster—and again, the gap widens as tasks get harder.

Frontend E2E testing

Once the first version of ego lite existed, our team got hooked on it fast. Faster than we expected, actually.

Tests run directly inside the browser where you're already logged in—no more spinning up test accounts or scripting a login flow first.

The bigger shift is how the agent reads the page. Instead of raw HTML, it gets a semantic snapshot straight from the Chromium engine—usually a couple hundred tokens for an entire page. All the noise—styles, scripts, and hidden nodes—is stripped out. What's left is just the buttons, inputs, and links that actually exist, along with their names.

Compared with Playwright MCP, we're roughly 2× faster in complex scenarios.

At this point, a lot of our frontend testing has become simple enough that non-engineers on the team can run it themselves. No scripts, no selectors—just describe the flow in plain language. Something like:

/ego-browser check whether this matches the expected behavior, run through the full flow, and report the results.

The agent opens the page in its own Space, drives it using your logged-in session, checks its progress against snapshots at every step, and reports the results and any errors it encountered. One sentence covers most cases—nobody has to write or maintain a test script.

We've also written up a bunch of other use cases on the site—scraping social posts, job hunting, apartment hunting, booking flights and hotels, tracking stocks, and that kind of thing. Worth a look:

How does this compare with what's already out there?

Before building ego lite, we tried the existing options ourselves. Each one had a clear gap:

First-party AI browsers (Comet, Atlas) only work with their own built-in agents. Whichever agent you already use—such as Codex or Claude Code—can't plug in, and there's no shared context, so these browsers never really fit into your existing workflow.
Chrome-bridging plugins and middleware (Agent-Browser, browser-use) can sometimes use your login state. But login-state inheritance is unreliable, windows and tabs pop open on their own, and the agent ends up fighting you for control of the same browser while it runs.
Agent-embedded browsers (the ones built into Codex and Cursor) are stripped down and incomplete. They can't inherit your login state, and they struggle with anything more complex than a simple page.

The common thread: none of them treat "a human and an agent sharing one browser" as a problem worth solving properly. That's the gap we're going after.

What's next

We're a small team moving fast, and ego lite is still rough around the edges. There's no point pretending otherwise. Here are the known limitations as of now:

Resource usage is on the heavier side. It's a full Chromium fork with multiple Spaces running in parallel, so memory usage adds up. We're actively optimizing it.
It's macOS-only for now. We wanted to nail one platform before spreading ourselves too thin. Windows requests have been piling up since launch, and it's next on the list.
There's no workflow reuse yet. The plan is to let successfully completed tasks turn into reusable skills, so agents don't have to figure out the same workflow from scratch every time.

About us

We're a small team that ships fast, and we're genuinely committed to open source. Our GitHub repository has been public for less than two months, yet it has already reached #1 on GitHub Trending and surpassed 5,200 stars. In a surreal moment yesterday, Twitter co-founder Jack Dorsey shared a screenshot after his new open-source project, Buzz, made the trending list—and there we were, ranked one spot above it.

Alongside ego lite, we've been contributing to Chromium itself. This year, our contributions have included the CSS shape() feature and improvements to animation frame rates during window resizing, which were mentioned at Google I/O 2026.

ego lite is free, with no account needed—just point your agent at it.

We know it's not there yet, and that's fine. We're planning to sharpen it release by release until it's a browser that genuinely works well for people and agents alike.

If you're already using Claude Code, Codex, or Cursor for web tasks, give it a try—and tell us what's broken. We want the real feedback.