What does wingman change about my agent's browser tools?

In the default hand-off mode it hides the browsing MCP server's page-action tools and routes them through one wingman tool, browse_step, while reading, tabs, dialogs, drag, screen-position clicks, waiting, and session tools stay with the agent.

How did it perform against Playwright alone?

Across 17 tasks run twice each way, wingman completed all 34 runs and Playwright alone 32 of 34, with both misses on the file-upload task; wingman spent $3.75 against $5.35 and was 0.5 s slower on the median.

Does it open its own browser?

No: it attaches to a real logged-in Chromium-family browser, finding Chrome, Edge, Brave, Chromium, Opera, or Vivaldi automatically in that order, so your sessions and cookies are already there.

How does the agent send work to Jev?

The agent sends its goal and steps through the browse_step tool; wingman reads the page, asks Jev to grade how well each element fits the step, acts, checks the result, and returns, handing the step back when unsure.

Can I keep some tools with the agent instead?

Yes: setting handoff.mode to optional hides nothing, and handoff.retain keeps named capability classes with the agent, such as script. Pointer-xy, drag, tabs, dialog, read, wait, and session always stay with the agent.

What does it need to run?

Node.js 22 or newer, a Chromium-family browser, and a TypeSafe API key in TYPESAFE_API_KEY or in a KEY=VALUE file named by secrets_file; install with npm install -g jev-browser-wingman.

Why does nothing change after I install it?

Because the top-level mode is off when absent: you must set 'mode' to 'on' in the config yourself. Only handoff.mode has a default, forced, which enables the hand-off behavior described above.

CommunityJev Agents & MCP

Jev Browser Wingman

Hands the agent's browser clicks and typing to Jev: wingman hides Playwright MCP's page-action tools, asks Jev which element fits each step, and returns the step to the agent when it is unsure.

See highlights

Open github.com

Stars: 7
Language: TypeScript
License: MIT
Published by: Community
Author: coderexpert123
Last verified: Oct 2026

coderexpert123/jev-browser-wingman

User repository on GitHub

View on GitHub

Browser automation where TypeSafe's Jev model picks each step. Attaches to an existing Chrome via CDP and runs alongside Playwright MCP. Not affiliated with TypeSafe.

Forks: 2
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity typescript mcp browser-automation playwright

Works forJev for Tool calling Jev for Decision gating

Used inJev for Software & DevTools

Languages: TypeScript
Platforms: MCP Playwright CLI
Capabilities: Open sourceRequires API key

Highlights

Attaches to an existing logged-in Chromium-family browser and drives clicks and typing through one tool, browse_step.
In the 68-run comparison it completed all 34 runs against Playwright alone's 32 of 34, at $3.75 versus $5.35 total.
The median run was 16.0 s versus 15.5 s, so wingman was 0.5 s slower on the median despite costing less.
Hand-off mode hides the browsing server's page-action tools and keeps tabs, dialogs, reading, drag, and waiting with the agent.
A step whose element it cannot pick goes back to the agent; Jev only grades how well each element fits the step.

Watch out

MIT and not affiliated with TypeSafe. The README says two runs per task make the speed and cost differences indicative rather than statistically established, and per-run cost swings with prompt-cache warmth. The top-level mode is off by default and decisions need a TypeSafe key.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

Hacker News: Show HN: Agent Chaperone screens tool calls and results with Jev before the agent sees them.

sepehrsafari5 pointsShow HN: Agent Chaperone – Screen AI agent tool calls and results with JevRead the thread on Hacker News

X: jev-review: a local-first score loop for coding agents

Reddit: One command to wire Jev into Claude Code, Claude Desktop, and Codex.

Related terms

Glossary definitions related to this entry.

MCP server

More Jev resources like this

BrowserClaw

Chrome MV3 extension plus a local MCP server exposing 50 browser tools; its chrome_act_toward_goal micro-loop uses TypeSafe Jev System One or a local decider to pick the next action on an 85%+ pruned, Shadow-DOM-pierced DOM tree, dispatching trusted CDP events in logged-in Chrome without stealing focus.

Jev Agents & MCP#community#mcp#browser-automation

Community

GoldenLoaf24h/browserpawTypeScript

agent-chaperone

An MCP proxy and hooks adapter that screens an agent's tool calls before they run and the tool results before the agent reads them, with thresholds in a policy file.

Jev Agents & MCP#community#mcp#security

Community

agent-chaperone/agent-chaperoneTypeScript

jev-code

Five typed tools — classify, check, score, rank, and ask — plus a skill, wired into Claude Code, Codex, Pi, and OpenCode in one setup command.

Jev Agents & MCP#community#typescript#mcp

Community

FrancoisChastel/jev-codeTypeScript

Public Browser

MCP server that lets Claude Code, Cursor and other clients drive Chrome over the DevTools Protocol, exposing an accessibility-tree menu of stable refs. Its example Jev loop has Jev pick browser actions from that menu, and the README reports about a third fewer tokens and a quarter fewer tool calls in a blind benchmark.

Jev Agents & MCP#community#mcp#browser-automation

Community

Silbercue/public-browserTypeScript

DSH Jev Plugin

A DeepSeek Harness plugin that exposes Jev's noul, choice and score primitives as agent tools, plus configurable guardrails, retries, response caching, a circuit breaker and Web UI result cards.

Jev Agents & MCP#community#typescript#deepseek

Community