jev-ra

Fast browser-use layer for coding agents: Jev picks the operation and target element each step, via MCP server or CLI.

MCP serverbetaMIT

Overview

jev-ra is a fast browser-use layer for CLI coding agents. Claude Code, Codex or any MCP client gives it a goal, and TypeSafe's Jev picks both the operation and the target element for every step in one round trip. The host agent plans, supplies text values, reads what the page says and takes over when jev-ra escalates; no second LLM runs inside the loop. It drives Chrome over CDP.

The README reports a benchmark on one machine through OpenRouter: Wikipedia navigation in 2.7 s versus 23.1 s for browser-use with Gemini 3 Flash, Google Flights in 8.9 s versus 66.4 s, and a Korean shopping-site sort in 3.8 s versus 15.1 s, a 3.96-8.50x speedup. Each jev-ra run (median of 5) was verified against the final page, 25 of 25 passing with no text-model calls; methodology, p90 and cost are documented. A site replays a recorded run. Installers set it up for Claude Code or Codex, and docs are available in Korean, Japanese and Chinese.

What the decision model decides

Type	Question	Answers
Choice	Which operation should be performed next?(inferred)	—
Choice	Which element is the target for this step?(inferred)	—

How it works

The coding agent hands jev-ra a goal and any text values
Each step, Jev picks operation and target element in one call
jev-ra executes via Chrome DevTools Protocol
On uncertainty, it escalates back to the agent
The final page is verified

Strengths

Reported 4-8x faster than browser-use on the tested tasks
No second LLM inside the loop
One-command install for Claude Code and Codex
Benchmarks with methodology published

Caveats

Benchmarks from one machine and three tasks
Requires an OpenRouter key
Text values must come from the host agent

Getting started

export OPENROUTER_API_KEY=sk-or-..., then uvx jev-ra install claude (or codex)

At a glance

Type: MCP server
Status: beta
Language: Python
Runtime: MCP server and CLI (Chrome CDP)
Has code: yes
Has tests: yes
License: MIT

Use cases and integrations

Agent Decisions340

How curated lists describe it

Browser use for coding agents, 3-5x faster than browser-use. MCP server + CLI; TypeSafe Jev decides every step in ~300 ms.
logicrw/awesome-jev-projects
Browser use for coding agents, 3-5x faster than browser-use. MCP server + CLI; TypeSafe Jev decides every step in ~300 ms.
hellogumbo/awesome-jev

Frequently asked questions

What does Jev decide?

The operation and target element for every browser step.

How much faster is it?

The README reports 3.96x to 8.50x faster than browser-use on three tasks.

Does it use an LLM internally?

No. The host agent supplies text values and takes over on escalation.

Related projects

jev-browse (0x7067)
Agent-plugin browser tool where Jev picks one operation and one row from a numbered page table; a small LLM only writes typed text.
ProjectbetaTypeScript
jev-browser-control
Chrome extension plus MCP server so Claude or Codex can drive your real Chrome, with Jev picking each click in about half a second.
MCP serverbetaTypeScript
jev-max
Jev browser agent that pierces shadow DOM and iframes, adds canvas fallback and full actions, with one Jev request per cycle.
MCP serverbetaPython
jev-mcp (legostin)
MCP server that lets Claude Code or Codex drive real Chrome with Jev picking elements and checking steps, plus safety gates and traces.
MCP serverbetaTypeScript
jev-ultrafast-mcp
MCP server that takes a whole browser task (URL, goal, checks) in one call and drives Chrome server-side with a decision model.
MCP serverbetaPython
jev-browser-skill
Readable reference skill for Claude Code and Codex where Jev picks from a numbered table of page controls, with an explainer site.
Agent skillprototypeJavaScript