Overview & result
jev-reflex is an MCP server that lets Claude hand a whole browser task to Jev. Jev makes each click, type and select decision, which the author puts at about 100 ms each, while Claude stays in charge of planning the task and verifying the result. The README's argument is that most browser steps are a choice from a list and do not need a round trip to a frontier model.
From the original post
Claude thinks, Jev reacts: an MCP server that hands browser tasks from Claude to TypeSafe's Jev (~100 ms per decision).
How Jev fits in the loop
- Claude calls browse with a URL, a goal that says when to stop and the exact strings to type.
- jev-reflex opens a background tab in your Chrome and reads the page into a numbered table of visible controls.
- Jev picks an operation and a target in one request: click, type, select, scroll, wait, done or blocked, each with a probability.
- Claude checks the result when the run ends.
How to reproduce
- Inspect the original showcase and source material at https://github.com/MSalvalaggio/jev-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
Claude hands over a whole browser task instead of one click at a time, so the frontier model is used for planning and verification only.
Limitations
- The roughly 100 ms per decision is the author's figure.
- Jev types only the strings Claude supplied; it never writes text.
- It drives a tab in your own Chrome.