wy-coliney/jev-browser-use

The core idea here is a division of labor: Jev, TypeSafe's browser-action model, handles clicks, scrolls, toggles, and navigation, while Codex retains responsibility for text input, visual interpretation, and final verification. The action loop stays inside an existing Computer Use connection, which means no new host-model turn fires for every click. That's the actual source of the speed claim—the README cites roughly 5–10× faster browser operations in the authors' own EZCollegeApp workflows, with the explicit caveat that server-side processing is excluded and the figure varies by task.

The cost argument is the sharpest part of the README. Jev 1.13 is priced at $0.042 per million input tokens with output listed as free. The comparison models—GPT-5.6 Terra at $2.00 input, GPT-5.6 Sol at $4.00, GPT-6 Astra at $10.00—run 48×, 95×, and 238× more expensive on input alone, before output charges. At 100,000 tokens, the difference is $0.0042 for Jev versus $0.20 to $1.00 for the others. The README is careful to note these are standard uncached rates checked September 18, 2026, and that total workflow cost also includes Codex.

The architecture is worth understanding clearly. Jev receives accessibility text, not screenshots. That's a meaningful constraint: it means Jev is operating on the DOM's exposed text layer, not doing vision-based inference. Codex handles anything requiring visual judgment or unsupported controls, and Jev resumes afterward. Sessions retain progress across handoffs and check fresh page state before acting.

Installation is a single npx skills add command for Codex, with a plugin path also available. The README is honest about prerequisites: Node.js 22+, Codex with Computer Use connected to Chrome or its in-app browser, and Jev access through TypeSafe or OpenRouter. Claude Code support is listed as coming soon; browser execution currently requires the Codex Computer Use runtime.

The troubleshooting section is unusually candid. It warns that a successful shell network request does not prove network access inside CUA, and that the browser tool is not listed inside functions.exec's deferred tools—a non-obvious gotcha that would otherwise produce confusing failures. The README also notes that page text goes to your selected provider, which matters for anyone delegating sensitive workflows.

This was built to solve a real, narrow problem: the authors were repeatedly clicking through college application evaluation screens during EZCollegeApp testing. The split-responsibility model generalizes to any agent workflow where repetitive navigation decisions are eating expensive model capacity. Whether the 5–10× figure holds outside their specific workflow is an open question the README doesn't oversell.