ego lite Setup: Let Claude Code Drive Your Real Browser

ego lite is a Chromium browser built so that you and your AI agent can use it at the same time without fighting over tabs, and it picked up roughly 986 stars today on its way to 3,779 total. The trick it sells is not automation, which every framework already does. It is login state: your agent works inside an isolated Space that inherits the cookies and sessions you are already signed into, so a task like "check my GitHub notifications" does not begin with a twenty-minute detour through a login wall. Setup runs about five minutes, and the whole thing is free.

• ego lite is macOS only right now; Windows and Linux sit on the public roadmap with no ship date.
• Installing the app registers an ego-browser command that any agent CLI can drive: Claude Code, Codex, Cursor, or your own.
• Each agent gets a fully isolated task space with its own tabs, so parallel tasks do not collide with each other or steal your cursor.
• The MIT license covers the skill in the repo, not the browser binary, which is a separate free download. Read that distinction carefully before you call this an open source browser.

What is ego lite and why is it trending?

The project comes from Citro Labs and landed on GitHub in April 2026. Three months later it is running near the top of the daily trending list, and the reason is a genuinely annoying gap that anyone who has pointed an agent at a real website has hit. Browser automation frameworks are libraries. They ship no browser, so they spin up a fresh profile that knows nothing about you, and half of every useful task becomes a login problem. AI browsers like ChatGPT Atlas and Perplexity Comet solve the login problem by being your browser, but the agent inside them is theirs, and the Claude Code session already open in your terminal cannot touch it.

RelatedSuperpowers Setup: Give Your AI Coding Agent a Real Workflow

ego lite splits the difference. It is one Chromium browser you use daily, and any agent CLI can drive it through a command called ego-browser. The isolation happens through what the project calls task spaces: separate browsing contexts, each with their own tabs, each inheriting your login state by default. Ten agents can work in ten spaces while your own window stays untouched. The maintainers benchmarked four automation tasks against Vercel's agent-browser and report finishing up to 2.5 times faster with substantially fewer tokens, which is a vendor number on a vendor-chosen task set, so weigh it accordingly.

How do you install ego lite on macOS?

Three paths, and they converge on the same place. The most direct is the DMG from the project's CDN, linked off the README for Apple Silicon and Intel separately: download, open, drag to Applications. If you would rather have your agent do it, the repo ships an installer script inside the skill that handles the architecture detection, the download, and the Gatekeeper quarantine flag:

# from inside the cloned repo, macOS only
$ sh skills/ego-browser/scripts/install.sh

The third path installs just the agent-facing skill, which is useful if you want the skill registered now and the app installed the first time an agent actually needs a browser:

# register the ego-browser skill with your agent
$ npx skills add citrolabs/ego-lite

However you get there, the app has to launch once. First run asks a single question: whether to migrate data from Chrome. Saying yes copies over logins, cookies, extensions and bookmarks, which is the entire point of the tool and also the moment to think about what you are handing an agent. Onboarding is what registers the ego-browser command, normally under ~/.local/bin, so skipping it leaves you with an app and no CLI.

How do you confirm the CLI is actually wired up?

Two commands settle it. First check the binary resolves:

$ command -v ego-browser

Nothing back usually means ~/.local/bin is missing from your PATH rather than a broken install. Patch it for the current shell and retry:

$ export PATH="$HOME/.local/bin:$PATH"
$ command -v ego-browser

Then prove the runtime works with the smallest possible script. Everything ego-browser runs is a Node.js heredoc, so the smoke test is one line of JavaScript:

$ ego-browser nodejs <<'EOF'
console.log('ego-browser ready')
EOF

Printing ego-browser ready means the harness is live and you can stop debugging.

How do you run your first agent browser task?

The interface is deliberately not a chat command. The agent writes JavaScript that calls preloaded helpers, and the whole multi-step sequence runs in one pass rather than in a call-look-call loop, which is where the token savings come from. A first task looks like this:

RelatedStrix Setup: Run an AI Penetration Tester on Your Code

$ ego-browser nodejs <<'EOF'
const task = await useOrCreateTaskSpace('inspect example page')
cliLog('task space id: ' + task.id)

await openOrReuseTab('https://example.com', { wait: true, timeout: 20 })

cliLog(await snapshotText())
EOF

The helpers cover the usual surface: click, fillInput, typeText, pressKey, captureScreenshot, waitForElement, scrollToBottomUntil, plus raw js and cdp escape hatches. The Node process exits after each heredoc and keeps no state, so every follow-up round starts by naming the same task space again. In practice you rarely type any of this: the skill teaches your agent the pattern, and you describe the task in English while it writes the script.

How does it compare with Browser Use, agent-browser and Atlas?

That last row is the one that decides it for most people. Browser Use and agent-browser run anywhere you can run Python or Node, and if your automation lives in CI or on a Linux box, ego lite is not in the conversation yet. What ego lite wins is the laptop case: a developer who wants their agent to act on sites they are personally signed into, right now, without building a credential pipeline first.

What are the gotchas before you rely on it?

Five, and the first two matter more than the rest. One, macOS only. The install script checks for Darwin and refuses everything else, and the roadmap lists Windows and Linux without dates. Two, the MIT license in the repo covers the ego-browser skill and its tooling. The README states plainly that the browser itself is a separate, free download, which means the actual Chromium build you are trusting with your cookie jar is not the thing you can read the source of. Free is not the same as open source, and the trending badge blurs that.

Three, Chrome migration is a genuine security decision, not a setup convenience. You are copying live sessions into a browser that an LLM will drive against pages that can contain instructions aimed at it. Prompt injection through page content is a real category, so keep an agent's spaces away from banking and admin consoles until you have watched it work. Four, the speed numbers are the maintainers' own, measured against one competitor on four tasks, and no third party has reproduced them. Five, this is a project three months old carrying 32 open issues and shipping betas weekly. The version numbering shows it: the tagged release sits at v1.2.5 from July 17, while the skill inside the repo already declares 1.2.6 dated July 20.

Our take

Most agent browser tooling has been solving the wrong half of the problem. Driving a page was never the hard part; whisper a CDP command and any Chromium obeys. The hard part is that the useful web is behind a login, and the standard answer has been to make developers rebuild their own authenticated session inside a sterile automation profile. ego lite skips that by making the agent a second tenant in the browser you already live in, which is either obviously correct or mildly alarming depending on how much you trust the agent, and honestly both readings are defensible.

The design detail worth stealing regardless of whether you install this is the interface shape. Handing the model a JavaScript runtime with preloaded helpers, instead of a menu of tool calls it invokes one at a time, lets it compose an entire multi-step task into a single script. That is a better fit for what these models are good at, and it is the part of the pitch that does not depend on any benchmark holding up. For today, if you are on a Mac and you already run Claude Code or Codex, this is a five-minute install with a real payoff. If you are on Windows or shipping automation to a server, bookmark the roadmap and check back.

Original analysis by GenZTech. Tool documentation: citrolabs/ego-lite on GitHub.