Jev Ego

Neither. The whole loop, TypeSafe plus the text helper plus CDP, runs inside a single ego-browser nodejs process, with no Python process and no Browser Harness hop. Reading a page for the first time after navigation costs 3 CDP calls, and no screenshots enter the agent loop because Jev consumes a structured element table. Guard, geometry and click resolve together in one Runtime.evaluate.

Jev only chooses the operation and its target from the numbered element table; target questions are speculative and only the one matching the chosen operation can execute. The text itself comes from a separate small text model behind TEXT_MODEL_API_KEY. Model output never becomes selectors, coordinates or executable JavaScript, so two decisions cost one network round trip.

The stated requirements are macOS with ego lite installed and ego-browser on PATH, developed against ego-browser 0.5.0.32, plus Node 20+, pnpm, TYPESAFE_API_KEY for decisions and TEXT_MODEL_API_KEY for typing. Tests are offline, but the live flights and smoke commands make paid API calls and drive ego lite. Nothing in the README documents a Linux path.

The 7.1 s Chrome figure belongs to upstream jev-ultrafast and the README explicitly says not to treat it as a jev-ego result. Measured on ego lite 0.5.0.32, flights finished in 31.6 s wall and 29.1 s decision clock with 11 actions, 15 TypeSafe decisions, 2 text-helper calls and 60 CDP calls. The stepper daemon kept observe at 0.11 s and act at 0.17-0.27 s.

macOS with ego lite installed and `ego-browser` on PATH (developed against 0.5.0.32), Node 20+ and pnpm. Set TYPESAFE_API_KEY for decisions and TEXT_MODEL_API_KEY for TYPE_TEXT, then run `pnpm install`, `pnpm build`, and copy `.env.example` to `.env`.

A stepper daemon plus one-shot commands: `serve` starts a loopback daemon on one TaskSpace, then `observe`, `act`, `suggest`, `step` and `run --goal` drive it, with `flights`, `smoke` and `probe` as demos. `stop` finishes the TaskSpace and removes the registry file.

MIT. The policy, snapshot and question text are derived from browser-use/jev-ultrafast, also MIT and by Browser Use, while the browser layer is ego lite. It is an adaptation of that policy rather than a from-scratch agent.