jet-browser Skill
当项目需要隔离的浏览器进程、有序的 JSONL 自动化、原生输入、截图或框架无关的浏览器工具时,安装、验证、集成或排查 Jet Browser WPE WebKit 运行时。不要用于控制已运行的个人 Chrome 配置;那种场景请使用环境的 CDP 工作流。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Jet Browser
Use Jet Browser as a replaceable browser-runtime boundary. Keep model choice, agent reasoning, credentials, and orchestration in the calling project.
Setup and verification
- Reuse an existing Jet Browser checkout when the project already contains one. Otherwise use
https://github.com/masakaai/jet-browserand keep the checkout location explicit. - Check for Docker and Node.js 24+ before running the standalone flow. If either is unavailable, report the missing requirement instead of claiming verification succeeded.
- Run
npm ci, thennpm run standalone. Treat its JSON result as the acceptance check: the local page, JavaScript marker, native text input, screenshot, network isolation, and clean shutdown must all pass. - For source changes, run
npm testandcargo test --all-targetsas well.
Integration
- Read
docs/demo.mdfor the container lifecycle anddocs/agent-tools.mdfor model-facing tool declarations. - Prefer the versioned declarations exported by
sdk/tools.mjs; do not invent tool names or response fields. - Send one JSON command per line and consume one JSON response per line. Preserve ordering and always close the session in guaranteed cleanup.
- Keep one isolated browser session per container. Mount profile or download paths explicitly and never pass unrelated host credentials into the container.
- Use semantic evidence for target selection and screenshots for visual verification when both are available.
- The runtime is WPE WebKit, not Chromium. Do not promise CDP, Chrome-extension, or Chromium-only compatibility.
Product boundaries
Jet Browser is local/self-hosted browser infrastructure, not an agent framework, anti-bot service, or hosted proxy network. For MASAKA's managed browser control plane, use the API documentation at https://masaka-ai.vercel.app/docs/ instead.
Benchmark claims must retain their scope. The published 1,445.25 ms ready time and 189.6 MiB active memory are medians from the pinned offline fixture; they do not measure public-site success, stealth, CAPTCHA handling, model accuracy, or long-running stability. See https://masaka-ai.vercel.app/jet-browser/tech-report/ for method and raw samples.