ego-lite: a macOS browser that hands your real logins to an AI agent

The tab fight ego lite is trying to end

The README frames the problem as contention rather than capability. Existing tools like browser-use and agent-browser are described as automation frameworks that need a separate browser to drive, where logins never carry cleanly and you and the agent end up fighting for the same tabs. That is a fair description of the usual arrangement: the agent launches its own Chromium profile, hits a login wall, and you either hand it credentials or script a session transfer. ego lite's answer is to be the browser you already use, with a second workspace bolted on for the agent. The target user is narrow and specific. You run an agent CLI such as Claude Code, Codex or Cursor in a terminal, you are on macOS, and the tasks you want automated sit behind a session you are already authenticated to. If your automation is a scheduled job against public pages, none of this applies to you.

Spaces, snapshots and the ego-browser skill

The architecture has three named parts. Spaces are isolated workspaces inside the single browser instance; the README says each agent or task gets its own, and that you can see which Space has an agent running and take it over or stop it. The Snapshot is the page representation the model reads, and the README claims kernel-level customization produces higher-quality snapshots, specifically calling out deeply nested iframes as a case where other approaches break down. That claim is not backed by a published method in the material provided, so treat it as a vendor statement rather than a measured result. The third part is ego-browser, the connection layer. It exposes browser capabilities as in-page JavaScript functions (the README lists snapshot, fill, click, wait, navigate and capture), and the agent writes a snippet that calls them in one pass. This is the design choice worth noting: the agent is expected to compose a multi-step task as code rather than issue a sequence of discrete commands. The README contrasts this with a CLI approach it characterizes as call two commands, look at the result, call two more, and states complex workflows finish up to 2.5x faster with fewer tool calls. No methodology, task set or hardware is given for that figure.

Installing it, and what the agent actually types

There are three install paths and they are not equivalent. The first is a .dmg for Apple Silicon or Intel from the CDN links in the README, which on install adds the ego-browser skill to every agent's skills directory on the machine. The second installs only the skill: npx skills add citrolabs/ego-lite, after which the first browser task walks you through installing the app. The third hands the job to the agent itself, by pasting a prompt that points at https://github.com/citrolabs/ego-lite and instructs it to read skills/ego-browser/references/install.md. That file is the authoritative setup document; the README does not reproduce its contents, so anything beyond the download and the npx line has to come from there. On first launch the app asks one question, whether to migrate Chrome data. Answering yes is what gives the agent your existing logins, cookies, extensions and bookmarks. Invocation at runtime is a slash command in the agent CLI: /ego-browser, a space, then a plain-language task, with the README's example being ego-browser follow @ego_agent on x.com for me. The README also states that browsing data stays on the device and that the only thing recorded is whether you opted into Chrome migration.

The Chrome migration is the whole bet, and the sharpest edge

Everything that makes ego lite useful flows from one setup decision: copying your Chrome profile. That is also where the risk concentrates. Copying cookies, extensions and bookmarks means the agent's Spaces inherit your authenticated identity across every site in that profile, not a scoped subset. The README does not describe per-site consent, a separate agent profile, or a way to grant access to one origin and withhold another. So the practical question is not whether the agent is trustworthy in the abstract but whether you want a process that can act as you on any site you are logged into. There is a second, quieter failure mode: the README says you and the agent work in parallel and that your tabs stay yours, but it also says you can take over a Space. Takeover implies shared state, and shared state implies that a half-finished agent action (a filled form, a cart, a draft) is something you may encounter. The documentation does not say what happens to a Space when you intervene mid-task, and that is worth testing before you point an agent at anything that spends money or sends messages. A third boundary is platform. ego lite runs on macOS today, with Windows and Linux listed on the roadmap. There is no headless mode described, which rules it out for CI pipelines.

Where it sits against the frameworks it names

The README's comparison table sets ego lite against browser-use, Vercel's agent-browser, ChatGPT Atlas and Perplexity Comet, and the honest read is that these are different categories. browser-use and agent-browser are libraries: you import them, you supply a browser, and you own the session problem. Their advantage is portability and control. They run on Linux, they run headless, they run in a container, and you decide exactly which profile or proxy they touch. ego lite trades all of that for the session problem being solved at install time. Atlas and Comet are the other direction: closed consumer browsers with their own built-in assistant, where the agent is the vendor's, not yours. ego lite's distinguishing position is that the agent is whichever CLI you already pay for, and the browser is a fork of your own data. If you need your automation to run on a build server at 3am, browser-use or agent-browser is the correct tool and ego lite is not, regardless of how the snapshot quality compares.

Licence, upgrades and the cost of staying current

ego lite is MIT-licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That is a permissive licence with no copyleft obligation, and it means you could fork the skill or the browser if the project's direction stops suiting you. It does not tell you anything about the hosted pieces: the CDN download URLs and the lite.ego.app homepage and docs are separate services, and the licence covers the repository. This is not legal advice; check the LICENSE file and your own obligations. On maintenance, the release cadence visible in the material is roughly monthly for stable tags (v1.2.3 in August, v2.0.0 in September 2026), with beta tags in between, and v2.0.0 landed the same day as the most recent push. That is a young project on a fast version curve, which means upgrade cost is real: a major bump between v1.2.3 and v2.0.0 suggests the skill interface or the app is still moving. Because the agent-facing surface is a skill installed into your agent's skills directory, an upgrade may require re-running npx skills add citrolabs/ego-lite or re-reading skills/ego-browser/references/install.md rather than just replacing the .dmg. The README also advertises an experience-accumulation feature that distills successful actions into reusable tools, marked coming soon, so that is not something to plan around yet.

Who this is built for, and what to check before you commit

The fit is a developer on macOS who already runs Claude Code or Codex in a terminal and whose browser work is gated by authentication: enriching leads, checking dashboards, pulling data from tools they are signed into. For that person, the install is short and the payoff is not having to rebuild a session. The misfit is anyone automating public pages on a schedule, anyone on Linux or Windows, and anyone whose security model requires the agent to hold credentials distinct from their own. One more thing to weigh: the headline numbers in the README (up to 2.5x faster on complex workflows, up to 5x on repeated tasks) come without a task set or measurement detail, and the 5x figure belongs to a feature listed as coming soon. Judge the project on the sharing model, which is concrete and verifiable, and treat the speed claims as unverified until you run your own tasks through it.

Editorial conclusion

Adopt ego lite if you are on macOS, already drive Claude Code or Codex from a terminal, and want an agent to work inside sessions you are already signed into. Do not adopt it if you need Linux or Windows (the README puts both on the roadmap), if you cannot accept a browser-level fork of your Chrome profile, or if your automation must run headless in CI. Before trusting it, verify three things yourself: that the Chrome migration is reversible, that the ego-browser skill resolves in your agent's skills directory, and that a Space can be stopped mid-task without leaving the shared profile in a modified state.

Official sources

1. citrolabs/ego-lite on GitHub
2. License: MIT
3. Project website
4. README
5. Releases