Jev Browser Use：Jev 负责点击导航，Codex 负责思考验收，浏览器自动化快 5-10 倍
Jev Browser Use: Jev Clicks and Navigates, Codex Thinks and Verifies — 5-10x Faster Browser Automation
📌 GitHub：https://github.com/wy-coliney/jev-browser-use Stars：33 | License：MIT | 语言：JavaScript 发布日期：2026-09-18
用大模型控制浏览器，有个一直没有解决好的问题：每次点击、滚动、跳转都要跑一次完整的模型推理。哪怕只是”点这个按钮”，也要截图、理解页面、思考、输出动作——这个开销太高了。
Jev Browser Use 的思路很简单：把动作拆开，按价值分配给不同的模型。
分工逻辑
目标 → 观察当前控件 → Jev 选择动作 → 执行 → 循环 → Codex 验收Jev（TypeSafe 决策模型）负责：
- 点击（Click）
- 页面跳转（Navigation）
- 标签切换（Tab switch）
- 滚动（Scroll）
- 开关切换（Toggle）
- 依据无障碍文本（accessibility text）判断目标控件，不依赖截图
Codex / Claude Code 负责：
- 输入文字（Typing）
- 读取页面内容并理解
- 做需要判断的决策
- 最终验收结果
这两类操作有本质区别。点击、滚动这类导航动作是结构化的决策：“这一堆控件里，下一步选哪个”——正好是 Jev 的强项（速度快、成本低、不需要生成文字）。而输入文字、理解页面语义、判断任务是否完成——这些需要真正的理解能力，留给 LLM 来做。
关键在于：动作循环全程在现有的 Computer Use 连接内进行，不新增额外的模型轮次。Jev 的每次决策在约 300ms 内完成，不打断当前任务的上下文。
安装
支持三种方式：
Codex Skill（推荐）：
npx skills add wy-coliney/jev-browser-use -g -a codex -yClaude Code Skill：
npx skills add wy-coliney/jev-browser-use -g -a claude-code -yCodex Plugin Marketplace：
codex plugin marketplace add wy-coliney/jev-browser-use手动安装：
git clone https://github.com/wy-coliney/jev-browser-use
node scripts/install.mjs依赖：Node.js 22+，已连接 Chrome 的 Codex Computer Use，以及 Jev 访问权限（TypeSafe 早期访问）。
使用方式
安装后，在任务描述里明确指定使用这个 skill：
Use Jev Browser Use on the settings page.
Open the filters, switch views, scroll through results,
and restore the original state. Independently verify the result.Jev 处理”打开过滤器、切换视图、滚动、还原状态”这些导航动作，Codex 在最后独立验证结果是否符合预期。
成本对比
以 10 万输入 token 为单位：
| 模型 | 成本 |
|---|---|
| Jev 1.13 | $0.0042 |
| GPT-5.6 Terra | $0.20 |
| GPT-6 Astra | $1.00 |
把高频的导航决策路由给 Jev，而不是全程跑 GPT-6 Astra，每 10 万 token 节省约 238 倍成本。这个差距在重复性自动化场景（比如批量处理表单、循环翻页、多次筛选）里会直接反映在账单上。
真实使用场景：EZCollegeApp 工作流
这个 Skill 最初是为 EZCollegeApp 的文书评估工作流构建的。典型流程包括：反复打开评估条目、展开备注、滚动报告、跳转到编辑器——这些操作重复量高、单步判断简单，正好适合批量路由给 Jev。
实测结果：相关工作流快了 5-10 倍。
设计背后的逻辑
Jev Browser Use 本质上是一个双层路由的实现：低认知、高频率的操作 → 快速便宜的决策模型；高认知、低频率的判断 → 强大的语言模型。
这和 TypeSafe Jev 本身的定位完全吻合——Jev 不适合”帮我写段文字”，但非常适合”从这 N 个选项里选一个下一步操作”。把浏览器里的每个动作理解成一道多选题，就能把 Jev 的并行决策能力用到实处。
随着浏览器自动化越来越常见（Agent 操作网页、填表、抓取、测试），这种分工模式可能会成为一个常见的工程范式：决策层用专用模型，推理层用通用模型。
开源代码与模型仅供学习、勿直接用于工作。
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。
📌 GitHub: https://github.com/wy-coliney/jev-browser-use Stars: 33 | License: MIT | Language: JavaScript Released: 2026-09-18
Using LLMs to control browsers has a long-standing inefficiency: every click, scroll, and navigation requires a full model inference pass. Even “click this button” means: screenshot, understand the page, reason, output action — an expensive loop.
Jev Browser Use’s premise is simple: split the work, assign each part to the model that fits it best.
The Division of Labor
Goal → observe controls → Jev chooses action → execute → repeat → Codex verifiesJev (TypeSafe’s decision model) handles:
- Click
- Navigation (page transitions)
- Tab switching
- Scrolling
- Toggle operations
- Reads from accessibility text — no screenshot needed
Codex / Claude Code handles:
- Text input and typing
- Reading and understanding page content
- Judgment calls and decisions
- Final verification
These two categories differ fundamentally. Navigation actions — “from this set of controls, which one is next?” — are structured decisions, exactly Jev’s strength (fast, cheap, no text generation required). Understanding page semantics, deciding when a task is complete, writing into fields — those need real comprehension, so they stay with the LLM.
Crucially: the action loop runs entirely within the existing Computer Use connection, adding zero extra model turns per click. Each Jev decision completes in roughly 300ms without interrupting the current task context.
Installation
Three options:
Codex Skill (recommended):
npx skills add wy-coliney/jev-browser-use -g -a codex -yClaude Code Skill:
npx skills add wy-coliney/jev-browser-use -g -a claude-code -yCodex Plugin Marketplace:
codex plugin marketplace add wy-coliney/jev-browser-useManual:
git clone https://github.com/wy-coliney/jev-browser-use
node scripts/install.mjsRequirements: Node.js 22+, Codex with Computer Use connected to Chrome, Jev access (TypeSafe early access).
Usage
After installation, specify the skill in your task description:
Use Jev Browser Use on the settings page.
Open the filters, switch views, scroll through results,
and restore the original state. Independently verify the result.Jev handles “open filters, switch views, scroll, restore state” as a navigation sequence. Codex independently verifies at the end that the result matches the goal.
Cost Comparison
Per 100K input tokens:
| Model | Cost |
|---|---|
| Jev 1.13 | $0.0042 |
| GPT-5.6 Terra | $0.20 |
| GPT-6 Astra | $1.00 |
Routing high-frequency navigation decisions to Jev instead of running GPT-6 Astra end-to-end cuts costs roughly 238x per 100K tokens — a difference that compounds quickly in repetitive automation tasks (batch form processing, paginated scraping, repeated filtering).
Real-World Use Case: EZCollegeApp
The skill was originally built for EZCollegeApp’s essay evaluation workflow. A typical session involves repeatedly opening evaluation entries, expanding notes, scrolling reports, and navigating to editors — high-repetition, low-judgment-per-step operations that route naturally to Jev.
Measured result: 5–10x faster on these workflows.
The Design Principle
Jev Browser Use is a two-tier routing implementation: low-cognition, high-frequency actions → fast, cheap decision model; high-cognition, low-frequency judgments → capable language model.
This aligns exactly with Jev’s positioning — it’s not for “write me a paragraph,” it’s for “pick the next action from N options.” Framing every browser interaction as a structured multiple-choice question lets Jev’s parallel decision architecture do real work.
As browser automation becomes a standard part of AI agent workflows (form filling, web scraping, testing, navigation), this routing pattern may become a common engineering baseline: decision layer uses a specialized model, reasoning layer uses a general-purpose one.
Open-source code is for learning purposes only — do not use directly in production work.
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18，不接受商业合作、不代表任何企业或机构立场，也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息，而是给出有独立思考的深入分析，希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型，仅供学习交流与技术借鉴。它们大多仍处于早期阶段，有待进一步研究和验证，请勿直接用于工作或生产环境；如需采用，请先自行充分测试，并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理，不代表文中提及的任何公司、产品、模型的官方立场，未与其构成商业关联或合作关系。
- 科技行业信息更新极快，我们尽力保证内容准确、及时，但不对完整性、实时性做绝对保证，具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有，我们会尽量注明来源；如你认为存在版权疑问或侵权，请通过下方邮箱联系我们，收到通知后会尽快核实处理（更正、加注来源或删除）。
- 文章内容仅为技术科普与个人观点，不构成投资、法律或其他专业建议，据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询：[email protected]
💬 评论与讨论
使用 GitHub 账号登录后发表评论