citrolabs/ego-lite

If you've wanted your coding agent to operate the sites you're already signed into, ego-lite is worth a look: it pairs a skill with a separate browser Space, but the browser itself is a closed download and only macOS and Windows are supported today.

The problem

Say you want Claude Code or Codex to follow a private account, pull data from a dashboard, or walk through a multi-step form on a site you're logged into. Handing the agent a raw Playwright session means replicating auth state, and a built-in agent browser competes for your tabs. With ego-lite you install a desktop app and the ego-browser skill; the agent then opens the task in its own Space, reads a structured text Snapshot of the page, acts, and reports back while your tabs stay untouched.

What's different

The core design choice is code-based control rather than CLI command loops: the browser's actions are exposed as JavaScript functions the agent calls directly, so it can compose a multi-step task into one script instead of alternating commands and observations. The author claims this cuts tool calls, tokens and completion time on complex tasks; the README benchmark (31 live-site tasks against Browser Harness, playwright-cli, chrome-devtools-cli and agent-browser) is the project's own measurement, not an independent one, so treat the 47–63% faster figure as vendor evidence.

The other notable mechanism is the Space: the agent gets an isolated browsing context but reuses your credentials. One early commenter called this exactly the setup they'd wanted, noting the agent gets its own independent space while still holding their credentials (HN, 2026-05-30). Setup is light: download the app, or add the skill via npx, which then walks you through installing the app.

The trade-off is that the automation runs against accounts you care about. A same-thread commenter asked how the browser masks itself against anti-bot detection and worried about getting an account flagged (HN, 2026-05-30); a reply said it worked for that person's own X scenario (2026-05-31), but this is a single anecdote, not a guarantee. If a site blocks automation, that risk lands on your account, not on a throwaway.

Who should try it

Good fit: a developer whose agent already uses Codex or Claude Code and who needs to automate logged-in sites around their own workday without surrendering their browser session. Also reasonable for people who want to run several agent sessions side by side; one commenter described using multiple agent-controlled sessions for browser games as a satisfying multiboxing setup (HN, 2026-05-30).

Poor fit: Linux users (on the roadmap, not available now), anyone who wants the browser source or a self-built binary, and people who need an off-the-shelf headless automation library rather than a desktop companion app. If your task is pure scraping on public pages, a normal headless setup avoids the account risk entirely.

Requirements and limitations

ego-lite ships as signed desktop downloads for Apple Silicon, Intel macOS and Windows x64; there is no Linux build. Installation can be done as a downloaded app or by adding the ego-browser skill with npx, after which the agent walks through app installation. The skill installs into every agent's skills directory on the machine, so it is a machine-level convenience rather than a per-project dependency.

The repository is MIT-licensed, but the browser itself is a separate free download, not open-source code in the repo; if you need auditable browser internals, that matters. Ongoing care is around credentials and site policies: keep an eye on which accounts the agent touches and whether a target site tolerates automated access.

HotRepo's take

HotRepo recorded the project at #3 on the weekly JavaScript board with +1,784 stars for the week of 2026-10-25, but has no comparable daily-board history, so persistence can't be judged from that single dated sample. Worth trying if you're on macOS or Windows and want logged-in automation from your existing agent; before relying on it, check the target site's policy on automated access, since the anti-detection question remains open in the project's own discussion.