PurrCat 把自己拆成六类插件，你想要什么，就换哪一类。下面逐类说清楚它能干什么、什么时候用得上——看完你就知道自己该开口要什么。 PurrCat splits itself into six kinds of plugins — whatever you want, one of them covers it. Here is what each one does and when to reach for it, so you know what to ask for.
界面 · UIInterface · UI
如果希望工作的流程更加可视化，希望固定的流程不用每次都重复用文字描述，就可以把它固化为界面插件——点点按钮即可表达自己的需求。 If you want your workflow to be more visual, and want a fixed flow you no longer have to spell out in words every time, turn it into a UI plugin — a few clicks then express what you need.
感官 · SensorSenses · Sensor
一个常驻的小进程，把外部消息推进会话，也能把你的回复送回原来的渠道。想把 PurrCat 接到飞书、邮件、RSS、闹钟这类外部世界时就用它。 A small long-running process that pushes outside events into a session and carries your replies back out. Reach for it when you want PurrCat wired into Feishu, mail, feeds, or an alarm.
编排 · GraphOrchestration · Graph
如果有一些比较固定的多 Agent 协作工作流、每一步都希望能得到人工审核，就可以用它，比如股票分析、每日文章总结等。 Reach for it when you have a fairly fixed multi-agent workflow where you want a human to review each step — stock analysis or a daily article digest, for example.
主循环 · AgentLoopMain loop · AgentLoop
声明每一轮循环在哪些阶段注入什么、检查什么，什么情况下不许收工；想定制自己的系统提示词也从这里下手。想给它定规矩、要求必须交付结果时就用它。 Declares what each loop stage injects and checks, and what counts as done. It is also where you write your own system prompt. Reach for it to set ground rules and demand a real result.
技能 · SkillKnow-how · Skill
写给它的说明书：语气、术语表、流程、参考资料，写成一份就随取随用；也可以直接从外部导入符合通用规范的技能。希望你每次写作或干活都照同一套规范时就用它。 A written brief for the agent: tone, glossary, process, references — write it once and it is always at hand, or import a ready-made skill that follows the common spec. Reach for it when you want the same house style every time.
外接工具 · MCPOutside tools · MCP
想给 Agent 提供外部工具能力，就用它；也可以从外部导入现成的 MCP 服务。 Use it to give the agent outside tool capability — and you can import ready-made MCP servers too.
有的时候，多个插件一起配合效果会更好。比如一条股票分析工作流，往往还会带上一个新造的数据获取工具——在上层插件里写好依赖关系，它们就会作为一个整体一起装上，效率一并提升。又再如说，sensor+skill 实现复杂社交媒体智能体，sensor 负责消息感知，skill 负责完成复杂的互动。 Sometimes several plugins work better together. A stock-analysis workflow, for instance, often comes with its own newly built data-fetching tool — declare the dependency on the upper-level plugin and they install as one unit, lifting their efficiency together. Or take a sensor plus a skill building a complex social-media agent: the sensor handles message awareness, while the skill carries out the intricate interactions.
