Stop writing prompts. Start designing loops.
Loop Engineering 的本地可视化控制驾驶舱 · Local visual control cockpit for Loop Engineering
Loop Cockpit 是一个本地运行的、面向独立开发者的 Loop Engineering 控制驾驶舱。
在你启动它的环境里，浏览器自动打开一个 Web 应用——你不再手写提示词，而是设计 Loop： 定义目标（Goal）、定义完成标准（Done Criteria），剩下的让 Agent 在沙箱里自主循环到达成为止。
2026 年 6 月，行业范式正在从 Prompt Engineering 转向 Loop Engineering。 但现有的工具要么是云端 SaaS、要么是命令行 Agent、要么是企业级运维平台—— 缺一个"独立开发者本地可视化设计 Loop"的工具。Loop Cockpit 来补这个空缺。
- 🎯 多 Loop 管理：一个工作台同时设计、调度、监控多个 Loop
- 🧰 Skill / MCP 配置：可视化挂载技能包与 MCP 服务器
- 💬 提示词控制：每个 Loop 可定义自己的 system prompt 模板
- ✅ 目标管理：Goal + Done Criteria 双向锚定，自动验收
- 📜 过程与结果可审计：全量日志、推理链、token 流水
- 🛡 沙箱运行：每个 Loop 在独立 Git Worktree 中执行，互不污染
- 📦 项目/任务双维度：可绑定指定 git 项目，可针对特定任务迭代
- ⏰ 多源触发器：Cron 定时、手动、Webhook、Git PR、Issue、邮件、飞书消息……
- 🔌 Agent 可插拔：MVP 优先 Claude Code，后续支持 OpenCode / Kimi Code / Codex / Trae / Qwen Code / MiniMax Code 等
🚧 Planning · Iteration 1 (Foundation) —— 正在做产品定义、原型、技术尖刀验证。 代码还没开始写，欢迎围观 路线图 与 PRD。
# 还没发布。等 Iteration 2 完成后会在这里放可运行命令。
npx loop-cockpit start # 计划中| 类目 | 文档 |
|---|---|
| 产品 | 白皮书 · PRD · 非目标 · 术语 · 路线图 |
| 设计 | 交互流程 · UI 规格 |
| 架构 | 总览 · 数据模型 · 模块 · ADR |
| 运行时 | Agent 契约 |
| AI 协作 | AGENTS.md |
这是一个独立开发者项目，目前由 @oh-summy 维护。
- 💡 想法和反馈：开 Issue
- 🐛 文档错漏：直接 PR
- 🤝 想协作：欢迎在 Discussions 留言
Loop Cockpit is a local-first, solo-developer-friendly Loop Engineering control cockpit.
It opens a browser-based web app on the machine you launch it from. Instead of writing prompts, you design Loops: define a Goal and Done Criteria, then let agents iterate autonomously in a sandbox until completion.
As of June 2026, the industry is shifting from Prompt Engineering to Loop Engineering. Existing tools are either cloud SaaS, command-line agents, or enterprise ops platforms — none of them is "a local, visual Loop designer for solo developers". Loop Cockpit fills this gap.
- 🎯 Multi-Loop management — design, schedule, and monitor multiple loops in one cockpit
- 🧰 Skill / MCP configuration — visually attach skill packs and MCP servers
- 💬 Prompt control — each Loop carries its own system prompt template
- ✅ Goal management — Goal + Done Criteria anchor the loop, auto-evaluation built-in
- 📜 Auditable process & results — full logs, reasoning trail, token accounting
- 🛡 Sandboxed execution — each Loop runs in an isolated Git worktree
- 📦 Project / task dual axis — bind a Loop to a Git repo, iterate per task
- ⏰ Multi-source triggers — Cron, manual, webhooks, Git PR / Issue events, email, Lark messages…
- 🔌 Pluggable agents — Claude Code first; OpenCode / Kimi Code / Codex / Trae / Qwen Code / MiniMax Code planned
🚧 Planning · Iteration 1 (Foundation) — product definition, prototype, and technical spike in progress. No runnable code yet. Track progress via the Roadmap.
# Not released yet. A working command will appear here after Iteration 2.
npx loop-cockpit start # planned| Category | Docs |
|---|---|
| Product | Whitepaper · PRD · Non-goals · Glossary · Roadmap |
| Design | UX flow · UI spec |
| Architecture | Overview · Data model · Modules · ADR |
| Runtime | Agent contract |
| AI collab | AGENTS.md |
This is an indie project maintained by @oh-summy.
- 💡 Ideas / feedback: open an issue
- 🐛 Docs typos: send a PR
- 🤝 Want to collaborate: drop a note in Discussions