Overview & result
solari-reflex is a speed layer for computer use on Solari browsers and Linux desktops. Each step is one structured observation, one Jev decision and one verified action, with no screenshots in the loop. The author reports a Stripe Checkout task finishing in 60.2 s against 194.9 s for the Codex CLI driving the same machines through Solari's MCP server.
From the original post
A speed layer for computer use on Solari: one structured observation, one Jev decision, one verified action per step. Browsers and Linux desktops.
How Jev fits in the loop
- Observe: one call reads the visible controls, values, text and a guard for each control, from the browser or the desktop accessibility tree.
- Decide: one Jev request picks the operation, a target under every operation, and whether the goal is done or blocked.
- Write: only for typing, a small model writes the value.
- Act: input goes to the observed control after a guard check, and the result is verified.
How to reproduce
- Inspect the original showcase and source material at https://github.com/hitakshiA/solari-reflex
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
Nothing the model returns becomes a selector, a coordinate or a script, and deny-listed controls are left out of the question entirely.
Limitations
- Timings are the author's measurements against the Codex CLI on the same Solari machines.
- It runs on Solari browsers and Linux desktops.
- Typing still uses a second, small model.