Tiller

A small model steering a real browser, one logit read at a time.

checking model

What should it do?

How it decides

Examples

click one to run it

Steps

  1. Each step, the page is reduced to a numbered list of visible controls. The model never writes an answer. It emits one token, and Tiller reads the probability of every option from that single pass.

    Busy pages run as a tournament: heats of 11 controls, then a final among the winners.

Test build. One shared browser, so runs take turns. Runs on your own sites are private: nobody else can watch them. Probabilities are relative to the options shown, not calibrated truth. Decision interface after SemIf, agent loop after jev-ultrafast.

Live view sandboxed Chromium, view only

The agent needs you

Which one should it pick?

It is not confident enough to choose alone. The candidates are outlined and numbered in the browser.

auto-picks in 45s

Create a free account

Already have an account?