[{"data":1,"prerenderedAt":174},["ShallowReactive",2],{"cms:global:en-US":3,"case-detail-en-qingyun-qyclaw-open-source-ai-workflow-platform":153,"cms:website-seo:20":170},{"basic":4,"theme":8,"seo":9,"footer":10,"social":15,"headerItems":19,"footerItems":58,"policyItems":146,"logoUrl":7},{"siteName":5,"tagline":6,"logoUrl":7,"faviconUrl":6},"广东倾云科技有限公司","","https:\u002F\u002Fwww.qingyunclouds.com\u002Fapi\u002Fpublic\u002Fassets\u002F1\u002Ffile",{"primaryColor":6,"accentColor":6},{"title":6,"description":6,"keywords":6,"canonical":6,"robots":6,"ogImageUrl":6},{"description":11,"heroDescription":11,"copyright":12,"contactHeading":13,"showLanguageSwitcher":14,"showBackToTop":14},"A technology R&D and service-oriented enterprise with the core of \"AI+Internet\", focusing on AI scenario transformation, software system construction and network infrastructure.","© CopyRight 2026 广东倾云科技有限公司 版权所有","Cooperation",true,{"website":16,"email":6,"phone":6,"address":17,"postalCode":18},"https:\u002F\u002Fwww.qingyunclouds.com\u002F","Room 717, 7th Floor, Torch Building, Zhongshan Port Street, Zhongshan City, Guangdong Province, China","508400",[20,28,34,40,46,52],{"id":21,"label":22,"kind":23,"openInNewWindow":24,"target":25,"children":27},"d6ba614a-7d89-4c43-a8ce-1d36866f59bf","Home","external_url",false,{"url":26},"\u002F",[],{"id":29,"label":30,"kind":23,"openInNewWindow":24,"target":31,"children":33},"nav-about","Abouts Us",{"url":32},"\u002Fabout",[],{"id":35,"label":36,"kind":23,"openInNewWindow":24,"target":37,"children":39},"nav-news","News",{"url":38},"\u002Fnews",[],{"id":41,"label":42,"kind":23,"openInNewWindow":24,"target":43,"children":45},"nav-cases","Case Center",{"url":44},"\u002Fcases",[],{"id":47,"label":48,"kind":23,"openInNewWindow":24,"target":49,"children":51},"nav-join","Careers",{"url":50},"\u002Fjoin",[],{"id":53,"label":54,"kind":23,"openInNewWindow":24,"target":55,"children":57},"nav-contact","Contact Us",{"url":56},"\u002Fcontact",[],[59,91,122],{"id":60,"label":61,"kind":62,"openInNewWindow":24,"target":63,"children":64},"footer-solutions","Solution","group",{},[65,71,76,81,86],{"id":66,"label":67,"kind":23,"openInNewWindow":24,"target":68,"children":70},"footer-retail","Smart Retail",{"url":69},"\u002F#solutions",[],{"id":72,"label":73,"kind":23,"openInNewWindow":24,"target":74,"children":75},"footer-office","Smart Office",{"url":69},[],{"id":77,"label":78,"kind":23,"openInNewWindow":24,"target":79,"children":80},"footer-city","Smart City",{"url":69},[],{"id":82,"label":83,"kind":23,"openInNewWindow":24,"target":84,"children":85},"footer-consumer","Intelligent C-end",{"url":69},[],{"id":87,"label":88,"kind":23,"openInNewWindow":24,"target":89,"children":90},"footer-tobacco","Digitalization of the tobacco industry",{"url":69},[],{"id":92,"label":93,"kind":62,"openInNewWindow":24,"target":94,"children":95},"footer-products","Products and Services",{},[96,101,106,111,117],{"id":97,"label":98,"kind":23,"openInNewWindow":24,"target":99,"children":100},"footer-llm","Development of Large Model Applications",{"url":69},[],{"id":102,"label":103,"kind":23,"openInNewWindow":24,"target":104,"children":105},"footer-ai","AI solution customization",{"url":69},[],{"id":107,"label":108,"kind":23,"openInNewWindow":24,"target":109,"children":110},"footer-software","Industry software customization",{"url":69},[],{"id":112,"label":113,"kind":23,"openInNewWindow":14,"target":114,"children":116},"footer-video","Video Creation Platform",{"url":115},"https:\u002F\u002Fwww.xuyingai.com\u002F",[],{"id":118,"label":119,"kind":23,"openInNewWindow":24,"target":120,"children":121},"footer-compute","Computing power server",{"url":69},[],{"id":123,"label":124,"kind":62,"openInNewWindow":24,"target":125,"children":126},"footer-company","Join us",{},[127,133,138,142],{"id":128,"label":129,"kind":23,"openInNewWindow":24,"target":130,"children":132},"footer-about","Company Profile",{"url":131},"\u002F#about",[],{"id":134,"label":135,"kind":23,"openInNewWindow":24,"target":136,"children":137},"footer-cases","Customer Case",{"url":44},[],{"id":139,"label":36,"kind":23,"openInNewWindow":24,"target":140,"children":141},"footer-news",{"url":38},[],{"id":143,"label":48,"kind":23,"openInNewWindow":24,"target":144,"children":145},"footer-join",{"url":50},[],[147],{"id":148,"label":149,"kind":23,"openInNewWindow":14,"target":150,"children":152},"5401443a-f196-4ff4-a970-5f6b79b6ae79","粤ICP备2021127448号",{"url":151},"https:\u002F\u002Fbeian.miit.gov.cn\u002F#\u002FIntegrated\u002Findex",[],{"id":154,"slug":155,"path":156,"title":157,"description":158,"date":159,"category":160,"cover":161,"client":6,"period":6,"tags":162,"metrics":163,"body":164,"seo":165,"relatedCases":169},6,"qingyun-qyclaw-open-source-ai-workflow-platform","\u002Fcases\u002Fqingyun-qyclaw-open-source-ai-workflow-platform","重磅开源！Qyclaw：国产多智能体操作系统，让AI协作像呼吸一样自然","倾云科技开源的企业级自动化 AI 工作流平台 —— Qyclaw。文章深入解析了基于 OpenClaw 架构打造的多 Agent 协同机制、飞书生态整合方案、流程编排引擎及企业级安全权限控制，全面展示了 Qyclaw 如何帮助企业快速搭建自主可控的 AI 自动化工作流。","2026-04-10T10:01:09.928+08:00","公司案例","https:\u002F\u002Fwww.qingyunclouds.com\u002Fapi\u002Fpublic\u002Fassets\u002F157\u002Ffile",[],[],"\u003Cp>Qyclaw 是一个面向多租户场景的智能体操作平台。\u003C\u002Fp>\u003Cp>它不是一个只会“聊天回复”的前端壳，也不是一个把整套智能体直接塞进容器里跑的黑盒。Qyclaw 的核心目标是：    \u003C\u002Fp>\u003Cul>\u003Cli>让智能体真正可运行、可隔离、可审计\u003C\u002Fli>\u003Cli>让用户记忆、技能、工具、外部连接都能分层管理\u003C\u002Fli>\u003Cli>让同一个平台可以在不同智能体后端之间切换\u003C\u002Fli>\u003Cli>让高风险执行进入容器沙箱，而不是把整套系统都绑死在容器里\u003C\u002Fli>\u003C\u002Ful>\u003Cp style=\"text-align: center\">\u003Cimg src=\"https:\u002F\u002Fwww.qingyunclouds.com\u002Fapi\u002Fpublic\u002Fassets\u002F157\u002Ffile\" alt=\"00.png\">\u003C\u002Fp>\u003Cp style=\"text-align: left\"> \u003Ca href=\"https:\u002F\u002Fgithub.com\u002F760485464\u002Fqyclaw.git\" rel=\"nofollow\">https:\u002F\u002Fgithub.com\u002F760485464\u002Fqyclaw.git\u003C\u002Fa> \u003C\u002Fp>\u003Ch2>项目定位\u003C\u002Fh2>\u003Ch2>Qyclaw 可以理解为一个“平台化的智能体操作系统”：\u003C\u002Fh2>\u003Cul>\u003Cli>上层是多用户、多会话、多技能、多连接器的平台能力\u003C\u002Fli>\u003Cli>中间是队列、调度、记忆、权限、审计等运行时编排能力\u003C\u002Fli>\u003Cli>下层是容器化工具执行沙箱\u003C\u002Fli>\u003C\u002Ful>\u003Cp>它适合这些场景：\u003C\u002Fp>\u003Cul>\u003Cli>团队内部的智能体工作台\u003C\u002Fli>\u003Cli>带审批和审计的自动化任务执行\u003C\u002Fli>\u003Cli>需要会话级工作区和文件处理的代理任务\u003C\u002Fli>\u003Cli>需要用户记忆、技能管理、MCP 接入的 AI 平台\u003C\u002Fli>\u003Cli>希望在 deepagents 和 claude 之间切换后端的系统\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>和 OpenClaw 相比，我们的优势\u003C\u002Fh2>\u003Ch2>如果把 OpenClaw 理解为一个偏“单智能体执行内核”的基础形态，那么 Qyclaw 的重点不是简单复刻，而是把它平台化、可运营化、可隔离化。\u003C\u002Fh2>\u003Cp style=\"text-align: center\">\u003Cimg src=\"https:\u002F\u002Fwww.qingyunclouds.com\u002Fapi\u002Fpublic\u002Fassets\u002F156\u002Ffile\" alt=\"01.png\">\u003C\u002Fp>\u003Cp>Qyclaw 的增强点主要在这几个方向：\u003C\u002Fp>\u003Ch3>1. 多租户，而不是单会话实验项目\u003C\u002Fh3>\u003Ch3>Qyclaw 原生支持：\u003C\u002Fh3>\u003Cul>\u003Cli>用户体系\u003C\u002Fli>\u003Cli>会话体系\u003C\u002Fli>\u003Cli>技能发布与审核\u003C\u002Fli>\u003Cli>会话级工作区\u003C\u002Fli>\u003Cli>多用户隔离\u003C\u002Fli>\u003Cli>管理端\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这意味着它不是“本地 agent demo”，而是可以向团队内部真实交付的平台。\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Ch3>2. 记忆是分层的，不是只有上下文拼接\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 现在把记忆拆成：\u003C\u002Fh3>\u003Cul>\u003Cli>用户长期记忆\u003C\u002Fli>\u003Cli>会话记忆\u003C\u002Fli>\u003Cli>记忆候选\u003C\u002Fli>\u003Cli>记忆审计日志\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这带来的好处是：\u003C\u002Fp>\u003Cul>\u003Cli>用户偏好可以跨会话延续\u003C\u002Fli>\u003Cli>当前会话状态不会污染所有会话\u003C\u002Fli>\u003Cli>记忆的读取、提取、更新都能审计\u003C\u002Fli>\u003Cli>后续可以继续扩展到 tenant 级记忆策略\u003C\u002Fli>\u003C\u002Ful>\u003Ch3>3. 工具、技能、连接器是分层建模的\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 不是把所有能力都混成“tool”。\u003C\u002Fh3>\u003Cp>当前系统已经区分：\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ccode>System Tool\u003C\u002Fcode>\u003C\u002Fli>\u003Cul>\u003Cli>平台内建能力\u003C\u002Fli>\u003Cli>比如 terminal、web_search、fetch_url\u003C\u002Fli>\u003C\u002Ful>\u003Cli>\u003Ccode>Skill\u003C\u002Fcode>\u003C\u002Fli>\u003Cul>\u003Cli>面向行为和流程封装\u003C\u002Fli>\u003Cli>支持草稿、发布、分组、共享、复制\u003C\u002Fli>\u003C\u002Ful>\u003Cli>\u003Ccode>MCP Connection\u003C\u002Fcode>\u003C\u002Fli>\u003Cul>\u003Cli>面向用户自己的外部能力接入\u003C\u002Fli>\u003Cli>比如 GitHub、Postgres、Notion、Custom HTTP\u003C\u002Fli>\u003C\u002Ful>\u003C\u002Ful>\u003Cp>这种拆分非常重要，因为它让权限边界更清楚：\u003C\u002Fp>\u003Cul>\u003Cli>平台能力归平台\u003C\u002Fli>\u003Cli>用户连接归用户\u003C\u002Fli>\u003Cli>行为封装归技能系统\u003C\u002Fli>\u003C\u002Ful>\u003Ch3>4. 容器化隔离的是“工具执行”，不是整套智能体运行\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>这是 Qyclaw 一个非常关键的设计边界。\u003C\u002Fh3>\u003Cp>Qyclaw 并不是把整个 agent runtime 扔进容器里跑，而是采用：\u003C\u002Fp>\u003Cul>\u003Cli>智能体编排在主系统\u003C\u002Fli>\u003Cli>工具执行在容器沙箱\u003C\u002Fli>\u003C\u002Ful>\u003Cp>也就是说：\u003C\u002Fp>\u003Cul>\u003Cli>队列\u003C\u002Fli>\u003Cli>会话状态\u003C\u002Fli>\u003Cli>记忆编译\u003C\u002Fli>\u003Cli>审批流\u003C\u002Fli>\u003Cli>SSE 推流\u003C\u002Fli>\u003Cli>多智能体后端切换\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这些都在主系统中完成。\u003C\u002Fp>\u003Cp>而下面这些高风险动作，才进入容器：\u003C\u002Fp>\u003Cul>\u003Cli>shell 命令执行\u003C\u002Fli>\u003Cli>文件读写\u003C\u002Fli>\u003Cli>技能脚本运行\u003C\u002Fli>\u003Cli>Office\u002FPDF\u002F文件处理\u003C\u002Fli>\u003Cli>工作区内工具调用\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这样做的优势是：\u003C\u002Fp>\u003Cul>\u003Cli>主系统更容易管理状态和恢复\u003C\u002Fli>\u003Cli>安全边界更清晰\u003C\u002Fli>\u003Cli>工具风险被限定在沙箱中\u003C\u002Fli>\u003Cli>不会因为容器生命周期把整套 agent state 搅乱\u003C\u002Fli>\u003C\u002Ful>\u003Ch3>5. 不止一个智能体后端\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 当前支持按会话切换执行后端：\u003C\u002Fh3>\u003Cul>\u003Cli>\u003Ccode>deepagents\u003C\u002Fcode>\u003C\u002Fli>\u003Cli>\u003Ccode>claude\u003C\u002Fcode>\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这意味着平台层和智能体后端是解耦的：\u003C\u002Fp>\u003Cul>\u003Cli>会话、记忆、技能、MCP、审计这些资产保留在平台里\u003C\u002Fli>\u003Cli>后端可以逐步切换、灰度、fallback\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这比“整个系统和单一 agent runtime 深绑定”更灵活。\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Ch3>6. 队列和调度能力更适合真实业务\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 不是简单的“一条消息直接跑一轮”。\u003C\u002Fh3>\u003Cp>系统已经引入：\u003C\u002Fp>\u003Cul>\u003Cli>会话级串行队列\u003C\u002Fli>\u003Cli>全局并发控制\u003C\u002Fli>\u003Cli>调度任务\u003C\u002Fli>\u003Cli>重试与退避\u003C\u002Fli>\u003Cli>backend 路由与 fallback\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这让它更适合：\u003C\u002Fp>\u003Cul>\u003Cli>多用户并发\u003C\u002Fli>\u003Cli>定时任务\u003C\u002Fli>\u003Cli>文件任务\u003C\u002Fli>\u003Cli>长任务执行\u003C\u002Fli>\u003Cli>人工审批恢复\u003C\u002Fli>\u003C\u002Ful>\u003Cp style=\"text-align: center\">\u003Cimg src=\"https:\u002F\u002Fwww.qingyunclouds.com\u002Fapi\u002Fpublic\u002Fassets\u002F155\u002Ffile\" alt=\"02.png\">\u003C\u002Fp>\u003Cp style=\"text-align: center\">\u003Cimg src=\"https:\u002F\u002Fwww.qingyunclouds.com\u002Fapi\u002Fpublic\u002Fassets\u002F154\u002Ffile\" alt=\"03.png\">\u003C\u002Fp>\u003Ch2>核心能力\u003C\u002Fh2>\u003Ch2>1. 多会话、多用户、多租户基础\u003C\u002Fh2>\u003Cul>\u003Cli>会话管理\u003C\u002Fli>\u003Cli>用户登录与权限体系\u003C\u002Fli>\u003Cli>会话消息流\u003C\u002Fli>\u003Cli>管理端\u003C\u002Fli>\u003Cli>会话级 backend 选择\u003C\u002Fli>\u003C\u002Ful>\u003Ch3>2. 智能体记忆\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Cul>\u003Cli>会话记忆\u003C\u002Fli>\u003Cli>用户长期记忆\u003C\u002Fli>\u003Cli>记忆候选提取\u003C\u002Fli>\u003Cli>记忆审计日志\u003C\u002Fli>\u003Cli>记忆与会话实例绑定\u003C\u002Fli>\u003C\u002Ful>\u003Cp>Qyclaw 的设计原则是：\u003C\u002Fp>\u003Cblockquote>用户长期记忆属于用户，由会话实例加载和更新； 会话记忆属于当前实例，不自动污染长期记忆。\u003C\u002Fblockquote>\u003Ch3>Hindsight 长期记忆集成\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 现已支持把用户长期记忆接入 Hindsight。\u003C\u002Fh3>\u003Cp>Hindsight 项目地址：\u003C\u002Fp>\u003Cul>\u003Cli>GitHub: https:\u002F\u002Fgithub.com\u002Flatent-knowledge\u002Fhindsight\u003C\u002Fli>\u003C\u002Ful>\u003Cp>当前接入方式如下：\u003C\u002Fp>\u003Cul>\u003Cli>长期记忆按 user:{user_id} 聚合，不再按会话拆 bank\u003C\u002Fli>\u003Cli>回答前调用 Hindsight recall，把与当前问题相关的用户长期记忆注入提示词\u003C\u002Fli>\u003Cli>回答后调用 Hindsight retain，把用户输入和助手回答写入长期记忆\u003C\u002Fli>\u003Cli>会话 conversation_id 仍保留在 metadata\u002Ftag 中，用于来源追踪，但不参与长期记忆隔离\u003C\u002Fli>\u003Cli>本地数据库仍保留聊天消息和会话摘要；长期语义记忆由 Hindsight 提供\u003C\u002Fli>\u003C\u002Ful>\u003Cp>当前默认不会在每轮主链路中同步执行 \u003Ccode>reflect\u003C\u002Fcode>。Qyclaw 采用低频异步反思策略：\u003C\u002Fp>\u003Cul>\u003Cli>每个用户累计完成若干轮 retain_turn 后，异步触发一次 Hindsight reflect\u003C\u002Fli>\u003Cli>触发阈值由配置项 hindsight.reflect_every_turns 控制\u003C\u002Fli>\u003Cli>设置为 0 时表示关闭自动 reflect\u003C\u002Fli>\u003C\u002Ful>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Ch3>3. 工具系统\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 目前对工具做了系统级建模：\u003C\u002Fh3>\u003Cul>\u003Cli>工具启用\u002F禁用\u003C\u002Fli>\u003Cli>风险等级\u003C\u002Fli>\u003Cli>是否需要审批\u003C\u002Fli>\u003Cli>是否必须在容器中执行\u003C\u002Fli>\u003Cli>backend 支持范围\u003C\u002Fli>\u003C\u002Ful>\u003Cp>高风险工具会受到更严格的限制，例如：\u003C\u002Fp>\u003Cul>\u003Cli>terminal 命令校验\u003C\u002Fli>\u003Cli>debug_exec 权限控制\u003C\u002Fli>\u003Cli>文件路径真实路径校验\u003C\u002Fli>\u003Cli>容器挂载范围收紧\u003C\u002Fli>\u003C\u002Ful>\u003Ch3>4. 技能系统\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 的技能系统不是简单的“prompt 模板”。\u003C\u002Fh3>\u003Cp>当前技能作用域包括：\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ccode>global\u003C\u002Fcode>\u003C\u002Fli>\u003Cli>\u003Ccode>group\u003C\u002Fcode>\u003C\u002Fli>\u003Cli>\u003Ccode>user\u003C\u002Fcode>\u003C\u002Fli>\u003Cli>\u003Ccode>conversation\u003C\u002Fcode>\u003C\u002Fli>\u003C\u002Ful>\u003Ch3>\u003C\u002Fh3>\u003Ch3>5. MCP 连接体系\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 把用户自己的外部能力建模为 MCP，而不是硬塞进全局 tool。\u003C\u002Fh3>\u003Cp>当前已经支持：\u003C\u002Fp>\u003Cul>\u003Cli>MCP server definition\u003C\u002Fli>\u003Cli>MCP connection\u003C\u002Fli>\u003Cli>MCP binding\u003C\u002Fli>\u003Cli>会话级绑定\u003C\u002Fli>\u003Cli>用户私有连接隔离\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这使得“用户自己的 GitHub \u002F Postgres \u002F Notion \u002F 自定义 HTTP 连接”可以被平台正式管理。\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Ch3>6. 定时任务和调度\u003C\u002Fh3>\u003Ch3>\u003C\u002Fh3>\u003Ch3>Qyclaw 已具备：\u003C\u002Fh3>\u003Cul>\u003Cli>定时任务模型\u003C\u002Fli>\u003Cli>任务运行日志\u003C\u002Fli>\u003Cli>scheduler\u003C\u002Fli>\u003Cli>任务统一进入 runtime queue\u003C\u002Fli>\u003C\u002Ful>\u003Cp>这意味着它不仅能做对话式 agent，也能做自动化 agent。\u003C\u002Fp>\u003Cblockquote>容器只做工具执行沙箱，不承载整套智能体运行。\u003C\u002Fblockquote>",{"metaTitle":166,"metaDescription":167,"canonical":168,"robots":-1,"ogImage":-1},"倾云科技 Qyclaw：开源企业级自动化 AI 工作流平台","推出开源企业级自动化 AI 工作流平台 Qyclaw。平台基于 OpenClaw 架构与飞书等办公生态深度融合，支持多 Agent 协同编排、工作流自动化及私有化部署，赋能企业高效构建智能工作流系统。","https:\u002F\u002Fwww.qingyunclouds.com\u002Fcases\u002Fqingyun-qyclaw-open-source-ai-workflow-platform",[],{"siteOrigin":171,"automaticCanonical":14,"managedOrigins":172},"https:\u002F\u002Fwww.qingyunclouds.com",[173],"https:\u002F\u002Fwww.gdqycloud.com",1794301315302]