Overview & result
Jev-RA lets a coding agent hand off a browser goal while Jev chooses the operation and target element for each step. The host agent supplies text values, reads page results, and takes over on escalation. Its README reports task-level benchmark timings; the ~300 ms per-step claim is not used here.
From the original post
A fast browser-use layer for CLI coding agents. Claude Code, Codex, or any MCP client hands jev-ra a goal. TypeSafe Jev, a System One decision model, picks the operation and the target element for every step in one round trip. Your agent plans, supplies the text values, reads what the page says, and takes over when jev-ra escalates. No second LLM runs inside the loop.
How Jev fits in the loop
- Capture page DOM snapshot and extract interactable candidate elements.
- Formulate the next step decision as a choice among actionable elements.
- Jev evaluates the candidate set against the immediate sub-goal.
- Chrome CDP executes the interaction and returns the resulting state to the host agent.
How to reproduce
- Clone brnyxx/jev-ra and follow its Chrome CDP setup.
- Define a targeted web exploration task (e.g. item lookup or link verification).
- Execute the reactive agent loop and track step execution timing and decision logs.
Why this build matters
Shows a concrete browser-agent decision boundary, with published task-level timing and success results that can be independently retested.
Reported performance
Reported by authorLatency: 2,714 ms median for the author's Wikipedia task (5 runs)
Limitations
- Requires reliable DOM selector generation from the underlying page state.
- Complex multi-page state transitions require an external planner to maintain macro goals.
- Network delays on heavy sites can outweigh decision latency.