pm-workbench
PM Workbench — AI 产品经理工作台
一句话定位:给产品经理的开源 AI 工作台——会议录音进来,原型/PRD/评审记录出去,Agent 有手、有记忆、还会自己造工具。
它解决什么问题
产品经理 80% 的时间耗在"搬运"上:开会记纪要、回来整理成需求、写 PRD、画原型、跟踪评审——每个环节都在重复输入。PM Workbench 把这条流水线交给一个 53 工具的 AI Agent:
会议录音(小程序实时转写)
→ AI 总结 + 自动提取需求改动点
→ 一键变更原型(Agent 直接改 HTML)
→ PRD 同步更新 → 评审记录自动留档
→ 需求全流程可追踪核心能力
能力 | 说明 |
🎙 会议闭环 | 小程序录音实时转写 → AI 总结 → 需求改动点提取 → 一键派原型 Agent 修改 |
🤖 53 工具 Agent | 主 Agent 编排 + 5 个专业子 Agent(原型/PRD/调研/评审/周报) |
🛠 自造工具(创造模式) | 跟 Agent 说"造一个工具",它自己创建并立即使用——借鉴 DeepSeek Harness |
🧠 越用越聪明 | Reflexion 教训回灌 + Voyager 成功打法库 + Mem0 式事实记忆 |
📋 六环 SOP | 一句"走完这个需求",调研→原型→PRD→评审自动流转 |
📖 MCP 协议 | 53 工具按标准 MCP 暴露,Claude Code / Cursor 可直连 |
💰 成本账本 | 每轮对话 token 成本追踪(DeepSeek 约 ¥0.01/轮) |
📱 全端覆盖 | Web 工作台 + 微信小程序(录音/零散点捕获) |
技术栈
Next.js 14 · TypeScript · DeepSeek(对话模型 + reasoner 双路由)· lowdb · 腾讯云 ASR · 微信小程序
快速开始
git clone https://github.com/huangshu-jing/pm-workbench
cd pm-workbench
npm install
npm run dev
# 打开 http://localhost:3000,设置页填入 DeepSeek API Key 即可架构(一句话版)
入口(Web / 小程序)→ Orchestrator 主 Agent(意图裁剪/分层提示词/上下文压缩)
→ 子 Agent 层(六环 SOP 流程引擎)
→ 能力层(53 工具 + 自造工具注册表)
→ 记忆层(事实抽取 + 向量检索)→ lowdb 数据层完整架构借鉴了 Pi Agent、OpenClaw、Mem0、MetaGPT、DeepSeek Harness 的设计,按垂类场景轻量化裁剪。
License
MIT
PM Workbench — AI Workbench for Product Managers
One-liner: An open-source AI workbench for PMs — meeting recordings in, prototypes/PRDs/review records out, with an agent that has hands, memory, and can even build its own tools.
PMs waste 80% of their time on "transporting" information: taking meeting notes, turning them into requirements, writing PRDs, drawing prototypes, tracking reviews. PM Workbench hands this pipeline to a 53-tool AI agent.
🎙 Meeting loop: record on your phone → live transcription → AI summary → extracted change points → one-click prototype edit by agent
🤖 Orchestrator + 5 specialized sub-agents (prototype/PRD/research/review/report)
🛠 Creation mode: say "build me a tool" and the agent creates and uses it immediately (inspired by DeepSeek Harness)
🧠 Gets smarter with use: Reflexion lesson injection + Voyager success-play library + Mem0-style fact memory
📋 Six-stage SOP: say "run this requirement end-to-end" and research→prototype→PRD→review flows automatically
📖 MCP protocol: all 53 tools exposed via standard MCP — connect from Claude Code or Cursor
Stack: Next.js 14 · TypeScript · DeepSeek (chat + reasoner dual routing) · lowdb · Tencent ASR · WeChat Mini Program
git clone https://github.com/huangshu-jing/pm-workbench && cd pm-workbench
npm install && npm run devMIT License
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/huangshu-jing/pm-workbench'
If you have feedback or need assistance with the MCP directory API, please join our Discord server