wy-coliney/jev-browser-use

What it solves

It reduces the latency and cost of browser-based AI agent workflows. By offloading repetitive navigation tasks—such as clicking, scrolling, and toggling—to a specialized, faster model (Jev), it prevents a high-capability LLM (like Codex) from having to process every single minor interaction, which typically slows down the process and increases token costs.

How it works

The project acts as a "Skill" for Codex, creating a hybrid execution loop. Jev handles the low-level browser controls by processing accessibility text rather than screenshots. When a complex action is required—such as typing text, interpreting visual elements, or performing a final verification of the result—the system hands control back to Codex. This allows the agent to maintain a session across handoffs without needing a new host-model turn for every single click.

Who it’s for

Developers and AI agent users who use Codex with Computer Use and need to automate repetitive browser operations (like navigating dashboards, reports, or settings pages) more efficiently and cheaply than using a general-purpose LLM for every step.

Highlights

• Increased Speed: Claims a 5–10× speedup in browser operations for specific workflows.
• Cost Efficiency: Uses Jev for repetitive decisions to significantly lower input token costs compared to frontier models.
• Hybrid Control: Splits labor between Jev (navigation/clicks) and Codex (text input/visual judgment/verification).
• No Extra Drivers: Works with existing browser connections without requiring additional npm dependencies or drivers.