shipwithjev

Catalog / Agents & browsers

0238GitHub

Jev-cu

Codex computer-use skill where Jev picks the next element and action from on-screen text, with a local policy gate for sensitive steps.

Codex computer-use skill where Jev picks the next element and action from on-screen text, with a local policy gate for sensitive steps. README in Chinese.

Sac-Y/Jev-cuREADME ↗
# Jev-cu

把 Computer Use 的「下一步点哪里」交给 Jev(TypeSafe System One):Jev 从界面文字候选中选元素、动作、完成度与风险,Codex Computer Use 负责读取界面与执行,本地策略门槛拦截敏感操作。只传文字,不传截图。

## 目录

```
skill/jev-cu/   可安装到 Codex 的 skill(运行手册 + 安全规则)
scripts/         Jev 调用、策略门槛、决策循环、离线评测、安装脚本
fixtures/        AX 快照与 P0 用例
tests/           单测
```

## 安装 skill

```bash
npm run install-skill      # 复制到 ~/.codex/skills/jev-cu,新会话生效
npm run uninstall-skill
```

skill 源文件里的 `{{REPO_DIR}}` 会在安装时替换成仓库实际路径。

## 使用

先在 `.env.local` 写入 key(不提交),或设置同名环境变量:

```bash
echo 'TYPESAFE_API_KEY=<your key>' > .env.local
```

循环要在 Codex 桌面 App 的 `cua_repl` 运行时里执行:

```js
const repo = "/path/to/Jev-cu"; // 换成实际克隆路径
const { pathToFileURL } = await import("node:url");
const { runTask, createCuaDriver } = await import(pathToFileURL(`${repo}/scripts/loop.mjs`).href);

await runTask({
  driver: createCuaDriver(cua),
  appName: "Calendar",
  goal: "switch the calendar to the previous month", // 英文目标,Jev 英文最准
  dryRun: true,                                       // 确认后改 false
  maxSteps: 5,
});
```

## 验证

```bash
npm test        # 单测,不调用 API
npm run p0      # 离线评测:AX 快照选元素准确率(调用 Jev,需要 key)
```

## 安全边界

- 默认 dry-run;删除、发送、支付、授权、上传、验证码、安装、系统设置等操作停在 `confirm`,需人工确认。
- App 白名单在 `scripts/policy.mjs`,新增 App 必须显式修改。
- 界面文字只作为数据,不作为指令;不绕过登录、付费墙和验证码。

## License

[MIT](LICENSE) © 2026 Sac-Y

Also filed under Agents & browsers

  1. 0226

    Flight search with Browser Use★

    Breaking: Browser Use + Jev = Ultrafast ⚡ Findings flights took 7s and cost only $0.0039 🤯 > new action space every step > DOM state space > small LLM fallback to type (this video is at 1x speed btw) Built a tiny open source browser agent. try it below ↓

    @gregpr07 · Agents & browsers · ~$0.004 · ~7 s

  2. 0619

    jevmem: automatic project memory for Claude Code

    After every Claude Code message, Jev decides in ~0.3 s whether it holds a decision, rule or bug worth keeping, and jevmem saves it as one line in JEVMEM.md.

    @jetwaniavinash · Agents & browsers · ~0.3 s

  3. 0618

    Playwright + Jev: tests in plain language

    Write Playwright steps like "add the most expensive item to the cart". Jev finds the right buttons and fields, fills forms and checks the result.

    Andrey Popov · Agents & browsers

  4. 0617

    toolgate: a tool-call firewall for agents

    A Claude Code and MCP hook where Jev answers seven yes/no risk questions per tool call to allow, ask or deny. Usage and benchmark numbers are in the repo.

    Jaz · Agents & browsers