中文版

paperclip（GitHub：https://github.com/paperclipai/paperclip ，2026-03-02 创建，今晨抓取时已 110,398★）是开源的 AI agent 团队编排应用，README 的定位是：“如果 OpenClaw 是员工，Paperclip 就是公司。”

它是一个 Node.js 服务器 + React 界面：自带 agent、定目标、组团队，从一个看板跟踪工作与成本；每个 agent 可选不同的模型与工具链，团队的任务、技能、权限与历史都放在一处。支持的 agent 包括 OpenClaw、Claude Code、Codex、Cursor、Gemini CLI、OpenCode、Pi、Hermes 等——README 的说法是“只要能接收心跳，就能入职”。四大支柱是：Agentic 任务管理（任务、审批与复核关卡）、Agent 组织架构（角色、权限与边界）、Agent 员工培训（Skill Studio、评测与绩效复核）、Agentic OS（跨厂商运行时、沙箱、集成与 MCP、SSO 与成本控制）。成本控制方面有公司、agent 与项目三级预算、阈值告警与到限自动暂停；治理方面有审批阶段、暂停/重派/停止与配置回滚。MIT 协议开源。

为什么值得了解

• 当 agent 从“一个助手”变成“二十个终端同时跑”时，缺的就不再是模型能力，而是管理：谁在干什么、花了多少钱、谁批了什么；Paperclip 把这套管理做成了一个像任务管理器一样的产品
• 对做 agent 产品、研究多 agent 协作的人是很好的参考实现：任务原子化签出、持久工作上下文、运行时技能注入、带回滚的治理，这几条设计可以直接借鉴
• 但它偏“公司形态”：个人只用一两个 agent 的场景未必需要这么重的组织架构，先想清楚自己的 agent 数量与协作复杂度再决定是否上

English Version

paperclip (GitHub: https://github.com/paperclipai/paperclip , created 2026-03-02, 110,398 stars at this morning’s fetch) is an open-source app for orchestrating teams of AI agents; the README’s positioning is: “If OpenClaw is an employee, Paperclip is the company.”

It is a Node.js server plus React UI: bring your own agents, assign goals, build the team, and track work and costs from one dashboard, with each agent able to use a different model and harness while the team’s tasks, skills, permissions and history stay in one place. Supported agents include OpenClaw, Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Pi, Hermes and more — per the README, “if it can receive a heartbeat, it’s hired.” Its four pillars are: agentic task management (tasks, approvals and review gates), an org chart for agents (roles, permissions and boundaries), agent employee training (Skill Studio, evals and performance reviews), and an agentic OS (cross-provider runtime, sandboxing, integrations and MCP, SSO and cost controls). Cost control covers company-, agent- and project-level budgets with threshold alerts and automatic pauses at configured limits; governance covers approval stages, pausing, reassigning or stopping work, and configuration rollback. MIT licensed.

Why it’s worth knowing

• When agents grow from “one assistant” to “twenty terminals running at once”, what you lack is no longer model capability but management: who is doing what, how much it costs, who approved what; Paperclip turns that management into a product that feels like a task manager
• For anyone building agent products or studying multi-agent collaboration, it is a strong reference implementation: atomic task checkout, persistent work context, runtime skill injection and governance with rollback are all designs worth borrowing directly
• But it leans toward the “company shape”: someone using just one or two agents may not need this much organisational machinery — decide based on how many agents you run and how complex their collaboration really is