Letting a frontier model plan, and a fast one click
Claude and Codex are good at deciding what to do. Clicking through a page is a different job, and sending every click through a frontier model is slow and expensive.
The challenge
When an agent drives a browser, the usual approach resends the whole conversation and the whole page to a large model at every step, just to choose the next button. Most of those steps are easy. The cost and the delay come from using the most capable model for the least demanding decision.
The system
- Claude plans
- Extension lists the page
- Jev picks the option
- Code maps it to the element
- Chrome acts
- Jev checks the goal
What we implemented
- Page lister
- The extension turns visible buttons, links and fields into numbered rows with their current values. Password fields are never listed.
- Jev, the step decision
- Answers which operation, which element, and whether the goal is already met, in one request of about half a second.
- Element mapper
- Code maps the chosen number back to the real element, so a model's answer can never become an invented selector.
- MCP server
- Exposes the browser to Claude, Codex and Grok Bot as tools, in your own Chrome with your own logins.
The result
Measured on the same live tasks, 27 to 198 times cheaper than letting Claude's or Codex's own model do the clicking, and more than twice as fast. Free and open source with your own key, or on prepaid credits.
Have a process like this one? Book a call.