Just tell the PM what you need, go grab a coffee, and come back to review the results.
你只需要跟 PM 说清楚要做什么,然后去喝杯咖啡,回来验收成果。
📜 Project Charter / 项目宪法 — CodeFlow = a lightweight AI Runtime / AI OS for multi-agent software development, driven by
@cursor/sdk, consumingfcop-mcp(consumer-side; does not define fcop). 码流 = 面向多 Agent 协作开发的轻量级 AI Runtime / AI OS,用@cursor/sdk驱动、应用fcop-mcp(消费方,不定义 fcop)。 Verbatim ADMIN charter quotes (5/9 10:48 + 10:51) live in design doc §0.0 / 完整原话见设计文档 §0.0🚀 CodeFlow v2 — repositioning to AI Runtime / AI OS ([draft]): English overview (5 min) · 中文速读 (5 分钟) · Full design doc / 完整设计文档
🟡 v1.0 alignment pending — The current backend kernel ships as
v0.1.0-rc.1(internal preview, not published to npm/PyPI). It implements CodeFlow protocol v0.1 (5 schemas: agent / task / review / session / skill) withReview.decision="needs_human"and ahuman_approvalsub-structure. These v0.1 concepts will be deprecated in v0.2 in favour of FCoP v1.0's Boundary capability — see FCoP issue #2 for the upstream v1.0 charter (7 abstractions, Boundary, etc.). CodeFlow v0.2 sprint 0 will fully align tofcop@>=1.0,<2.0. Release notes:docs/releases/v0.1.0-rc.1.md.
We published the methodology — How to Build an Automated AI Development Team in Cursor — and then built the tool to make it real.
我们先发布了方法论 — 如何在 Cursor 中搭建 AI 自动化开发团队 — 然后把它做成了产品。
CodeFlow (码流) is the production-ready tool born from that methodology. It turns the "filename as protocol" concept into a complete human-AI collaboration system: mobile command center + PC execution engine + multi-agent coordination.
码流(CodeFlow) 就是这套方法论的产品化落地。它把"文件名即协议"的理念变成了完整的人机协作系统:手机主控台 + PC 执行机 + 多角色自动调度。
Scope / 范围:FCoP vs this repo — The FCoP protocol and PyPI packages (fcop, fcop-mcp) are maintained in a separate GitHub repository (joinwell52-AI/FCoP). This repository is the CodeFlow / Bridgeflow tool line (Desktop, PWA, relay, etc.); it is not the FCoP source home. Boundary (中文) · README.zh — 本仓与 FCoP · README.en — This repo vs FCoP.
| OCR (legacy) | CDP — CodeFlow | |
|---|---|---|
| Latency 延迟 | 300–800ms | 10–15ms |
| Accuracy 精度 | ~90% | 100% |
| Agent Detection 角色识别 | Screenshot guess | DOM exact match |
| Setup 配置 | Manual shortcut edits | Zero — auto-injected |
| Fallback 降级 | None | Auto-falls back to OCR |
Zero-config CDP injection: CodeFlow detects on startup whether Cursor has the debug port open. If not, it silently restarts Cursor with
--remote-debugging-port=5253— no manual shortcut edits, no config files, nothing.零配置 CDP 注入:Desktop 启动时自动检测 Cursor 是否开启调试端口,若未开启则静默重启并自动注入参数,用户无需手动改快捷方式或任何配置文件。
| Methodology 方法论 | Product 产品 | |
|---|---|---|
| Repo | joinwell52 | codeflow-pwa |
| What | How to name agents, define roles, route tasks via filenames | Desktop EXE + PWA + Relay + MCP Plugin |
| Core idea | TASK-date-seq-Sender-to-Recipient.md |
Same protocol, automated end-to-end |
| Roles | PM / DEV / QA / OPS | 3 team templates (dev / media / mvp) |
| Human role | Tell PM what to do | Phone sends task, PC executes |
Desktop — download EXE (~35MB) and double-click:
- China: https://gitee.com/joinwell52/cursor-ai/releases
- GitHub: https://github.com/joinwell52-AI/codeflow-pwa/releases
PWA — open on phone and add to home screen:
For full documentation, see English | 中文.
Want to build from source or contribute? See CONTRIBUTING.md.
# Desktop (Python 3.12, Windows)
cd codeflow-desktop
pip install pyautogui pyperclip pywin32 websockets winocr Pillow watchdog psutil
python main.py
# PWA (no build step)
python -m http.server 8080 -d web/pwaProject structure:
codeflow-desktop/ # Desktop app (Python → EXE)
codeflow-plugin/ # Cursor MCP plugin
web/pwa/ # PWA (HTML/JS, no framework)
docs/ # Bilingual documentation
server/relay/ # WebSocket relay
- FCoP — File-based Coordination Protocol — the distilled protocol spec and essays behind CodeFlow's
.cursor/rules/design / CodeFlow.cursor/rules/背后的协议规范与随笔集 - When AI organizes its own work — a 4-AI-role team delivering 87 person-days in 17 days / 4 个 AI 角色在 17 天交付 87 人日
- An unexplainable thing I saw: the agent didn't just comply with rules — it endorsed them — an agent spontaneously wrote itself 4 FCoP memos on an unrelated task. 中文 · CSDN · Dev.to · Cursor Forum
MIT License. © 2026 joinwell52-AI
- Methodology: joinwell52-ai.github.io/joinwell52
- Product: github.com/joinwell52-AI/codeflow-pwa
- Changelog: CHANGELOG.md

















