citrolabs/ego-lite

The core architectural bet here is that browser automation fails not because agents are bad at clicking, but because the tool-call loop is expensive. Every conventional approach — browser-use, Vercel's agent-browser — works by exposing CLI-style commands the agent calls one at a time, inspects the result, then calls again. ego lite breaks that loop by wrapping browser capabilities as JavaScript functions the agent composes into a single output. Instead of a back-and-forth negotiation with the page, the agent writes one snippet that calls snapshot, fill, click, navigate, and wait in sequence, and ego-browser executes it in one pass. The README claims this finishes complex workflows up to 2.5× faster with fewer tool calls, benchmarked against Vercel's agent-browser across four tasks.

The other structural difference is the Space model. Most automation tools either hijack your existing browser or spin up a headless one you never see. ego lite runs as your daily browser, with each agent task isolated in its own Space — a fully separated workspace inside the same application. You keep your tabs; the agent keeps its own. Multiple agents can run in parallel Spaces simultaneously without interfering with each other or with your mouse. That's a real usability distinction, not a cosmetic one.

Login inheritance is handled by optionally migrating your Chrome profile on first launch — cookies, extensions, bookmarks, the lot. This is the part that automation frameworks consistently fumble. ChatGPT Atlas and Perplexity Comet also inherit Chrome data, but those products lock you into their built-in agent. ego lite accepts any external agent CLI: Claude Code, Codex, Cursor, or a custom one, all routed through the ego-browser skill layer.

The page snapshot quality claim is harder to evaluate from the README alone. The argument is that kernel-level customization lets ego lite handle deeply nested iframes reliably, where other tools break. That's a specific, testable claim, and it's the kind of thing that matters enormously in practice — a snapshot that misreads the DOM structure sends the agent down the wrong path entirely.

The experience accumulation feature, which would distill successful actions into reusable workflows and promises up to 5× speedups on repeated tasks, is listed as coming soon. That's the most interesting long-term proposition in the whole README, and it's not here yet.

The macOS-only constraint is real and current. Windows and Linux are on the roadmap but not shipped. For anyone building agent workflows on Linux infrastructure, this is a blocker today, not a footnote.