Cheap model clicks, smart model checks

What it does

Jev Browser Use is a Codex Skill that splits browser automation into two roles: Jev does the mechanical work — navigation, clicks, toggles, scrolling — while Codex keeps the judgment calls: text input, visual interpretation, and verifying the result. It rides your existing Computer Use connection to Chrome, with no extra driver or npm dependencies. The authors report roughly 5–10× faster browser operations in their own workflows, though they’re upfront that this varies by task.

The interesting bit

The action loop stays inside the existing Computer Use session, so every click doesn’t trigger a fresh host-model turn — that’s where the speedup actually comes from. Jev works from accessibility text rather than screenshots, which is presumably why it’s so cheap: $0.042 per million input tokens, with free output, versus dollars for the big models.

Key highlights

• Division of labor: Jev acts, Codex types, judges, and verifies; Codex can take over mid-task and hand back
• Sessions retain progress across handoffs and re-check page state before acting
• Works with TypeSafe or OpenRouter keys (not interchangeable), plus a native Codex plugin option
• Built for real drudgery: reviewing essay evaluations in the authors’ EZCollegeApp product
• MIT licensed, explicitly independent of OpenAI and TypeSafe

Caveats

• Browser execution currently requires the Codex Computer Use runtime — Claude Code support is “coming soon”
• The 5–10× figure is the authors’ own approximation, server-side processing excluded
• Workflows were exercised on macOS; page text goes to your chosen provider, so mind what you delegate

Verdict

Worth a look if your Codex agent burns tokens clicking through dashboards and settings pages all day. If you don’t already run Codex with Computer Use, there’s nothing here for you yet.