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.
Overview
jev-mcp by legostin is an MCP server that lets Claude Code, Codex and other MCP agents drive a real Chrome browser. The host agent stays the planner; jev-mcp turns the page into a compact, filtered model and Jev makes the many small decisions: which element is the departure-city input, which autocomplete suggestion to pick, whether a cookie wall blocks the page, and whether a step succeeded.
Each answer carries a confidence, so the tool acts on its own when sure and asks the agent (human-in-the-loop questions) when not. Every decision is traced and can be replayed. The README compares this with frontier LLM browser agents, citing about 100 ms per decision and $0.042 per million input tokens, and reports a measured run on a local airline-search fixture with a consent wall, autocomplete and pagination.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which element is the departure-city input? | — |
| Choice | Which autocomplete suggestion matches the intended value?(inferred) | — |
| Yes / no | Does a cookie wall block the page? | yes · no |
| Yes / no | Did that click work? | yes · no |
How it works
- The host agent plans and calls jev-mcp tools
- jev-mcp builds a compact model of the live Chrome page
- Jev answers element, suggestion and success questions
- High confidence acts; low confidence asks the agent
- Every decision is written to a replayable trace
Strengths
- Keeps HTML and screenshots out of the agent's context
- Confidence-driven act, verify or ask behaviour
- Replayable debug traces
- Safety gates on actions
Caveats
- Requires Node 22.18+ and a TypeSafe key
- Measured results come from a local fixture site
- Cost and latency figures are the README's own comparison
At a glance
- Type
- MCP server
- Status
- beta
- Language
- TypeScript
- Runtime
- Node MCP server controlling Chrome
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
MCP server that lets Claude Code, Codex & AI agents drive a real Chrome browser with JEV (TypeSafe System One): ~100 ms typed decisions, calibrated confidence, human-in-the-loop questions, safety gates and a debug trace.
logicrw/awesome-jev-projectsMCP server that lets Claude Code, Codex & AI agents drive a real Chrome browser with JEV (TypeSafe System One): ~100 ms typed decisions, calibrated confidence, human-in-the-loop questions, safety gates and a debug trace.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide here?
Small browser decisions such as which element to use, which suggestion to pick, whether an overlay blocks the page and whether a step worked.
Who does the planning?
The host agent, such as Claude or Codex; Jev does not plan or write text.
What happens when Jev is unsure?
The tool asks the agent instead of acting.
Related projects
Agent-plugin browser tool where Jev picks one operation and one row from a numbered page table; a small LLM only writes typed text.
ProjectbetaTypeScriptChrome extension plus MCP server so Claude or Codex can drive your real Chrome, with Jev picking each click in about half a second.
MCP serverbetaTypeScriptLocal gateway and MCP tools that let Jev predict which tool a coding agent should call next, from suggestions to enforced choices.
MCP serverbetaTypeScriptJev browser agent that pierces shadow DOM and iframes, adds canvas fallback and full actions, with one Jev request per cycle.
MCP serverbetaPythonFast browser-use layer for coding agents: Jev picks the operation and target element each step, via MCP server or CLI.
MCP serverbetaPythonMCP server that takes a whole browser task (URL, goal, checks) in one call and drives Chrome server-side with a decision model.
MCP serverbetaPython