Ego Jev — ZephyrDeng

No. It drives the ego-browser TaskSpace and Page API with your real browser, sessions, and logins. For a standalone agent with the same loop on other browsers, the digest points to jev-ultrafast instead.

Pass values in and the loop fills only matching fields by your hints; fills with no matching value escalate instead of inventing data. Pay, delete, upload, confirm, login pages, free text, low confidence, and repeats also escalate.

Install with the skills command, put one backend key in secrets.env, run preflightJev before creating TaskSpace or navigating, then call runJevLoop with goal, values, and verify. Missing credentials pause before any browser work.

About 300-550 ms and 11K input tokens per step on the direct backend. done is only a claim your verify callback decides, and opt-in record adds per-step requests plus a cookie diff with headers, bodies, and values off by default.

npx skills add ZephyrDeng/ego-jev, or the Claude Code plugin, or gh skill install; it is MIT licensed. Put one key in ~/.config/ego-jev/secrets.env: either TYPESAFE_API_KEY, the fastest path at about 300-550 ms per step with a $5 monthly credit and no card, or AI_GATEWAY_API_KEY to reach Jev through the Vercel AI Gateway.

On a 7-step form task run three times, all verified at about 5.5 s wall, Jev decided in a median ~270 ms. One bare LLM call over the same snapshot took about 1.0 s on Haiku 4.5, 1.4 s on Sonnet 4.5 and 2.8 s on GPT-5 mini, before any agent overhead. The comparison is against a single model call, not a full agent turn.

Numbered dN entries cover div cards, cursor:pointer widgets, and iframe and shadow content that never receive accessibility refs, which the README says are exactly the clickables that usually make a loop give up. Preflight also checks key resolvability before any browser work, though it does not confirm that the key or the service is actually live.