ego lite: a browser your AI agents can share with you

The hard part of letting an AI agent work on the web for you is rarely the model — it is the browser. Spin up a blank one and every login is gone, so the agent re-enters passwords and solves CAPTCHAs; connect it to your daily Chrome and it starts stealing tabs and the mouse while you are reading. ego lite, from Citro Labs, is a Chromium browser designed from the start for you and your agents to share: you keep browsing in front, agents work in their own Spaces. The repository was created in April 2026 and has climbed past 19,000 stars; it is MIT-licensed, the browser itself is a free download, and it runs on macOS and Windows (a Windows beta arrived on October 20), with Linux on the roadmap.

Core features

• A Space for every agent: each agent gets an isolated workspace where it opens pages, clicks and fills forms; several agents run in parallel without colliding, and you can jump into any Space to watch, take over, or stop it.
• Logins carried over intact: first launch asks exactly one question — whether to migrate your Chrome data. Bookmarks, passwords, extensions and cookies move over, so the agent opens Gmail, X or your company intranet already signed in, past the login wall that stalls blank browsers.
• Snapshot instead of raw HTML: inside the browser engine, ego lite compresses the page into a structured semantic tree where every button and input carries a stable number (@N). The agent locates elements by number, so a site renaming its classes does not break it, and deeply nested iframes and Shadow DOM are covered.
• Code-based automation: instead of “click, ask the model, click again”, the agent writes one JavaScript snippet calling the exposed tools — snapshot, fill, click, wait, navigate, capture — and runs it in a single pass. On the official comparison the batched approach used about 1.1k tokens where step-by-step calls used 13.6k.
• An open connection layer: ego-browser is the bridge between the browser and any agent CLI — Claude Code, Codex, Cursor, Gemini CLI, OpenCode — and it runs on your existing agent’s model, so there is no separate API key. Browsing data stays on your device.

The vendor also publishes Real-World Bench: 31 tasks on live websites, run by the same agent (pi) on the same model (gpt-5.6-sol, thinking effort max) and graded by the same judge. Reported results: 93.5% perfect completion versus 77.4% for the best of the other four tools (Browser Harness, playwright-cli, chrome-devtools-cli, agent-browser), an average of 6 minutes 38 seconds per task — about 55% faster than Browser Harness — at $1.75 versus $3.14 per completed task, a 47–63% speed lead across the field. It is a vendor-run benchmark, but the harness and raw results are open, so you can rerun it.

Typical use cases

• Signed-in web work: checking an intranet admin, triaging mail, managing social accounts — the pages a blank browser cannot enter are exactly where inheriting your logins pays off.
• Parallel multitasking: one Space researches, another tests the admin panel you just wrote, a third sweeps competitor sites; nobody hijacks anybody’s tabs and your mouse stays where you left it.
• Automation outside CI: writing deterministic product test suites remains Playwright’s job — ego lite addresses the other problem, having an agent do web chores for you. For natural-language web and mobile testing, see e2e.
• Boundaries worth knowing: an inherited login state is real power, so irreversible actions — payments, deletions, sending messages — deserve your own confirmation; Windows is a fresh beta; and the promised “experience accumulation that makes repeat tasks up to 5x faster” is still marked coming soon.

Quick start

1. Download the app from the GitHub repository or the website (macOS / Windows); installing it adds the ego-browser skill to every agent’s skills directory on your machine.
2. To install only the skill: npx skills add citrolabs/ego-lite — or paste the repository URL into your agent and let it follow skills/ego-browser/references/install.md by itself.
3. In your agent CLI, type /ego-browser followed by a plain-language task, for example “open my admin panel, exercise query, create and edit, and log every error”. Choose to migrate Chrome data on first launch to inherit your logins. If you would rather drive your own browser with Playwright, see browser-control.

Summary

ego lite’s value is not “one more browser” but a considered answer to what a browser shared by a human and their agents should look like: Spaces for parallelism and control, login inheritance for the real world’s login walls, Snapshot and batched code for speed and cost. Whether the 93.5% holds up outside the vendor’s benchmark is something individual tasks will tell — but the direction is clear: browsers were designed for one human and one mouse, and that assumption is being rewritten.