"The factory of the future will have only two employees: a man and a dog. The man will be there to feed the dog. The dog will be there to keep the man from touching the equipment." — Warren Bennis
「未來的工廠只有兩名員工:一個人和一條狗。人負責餵狗,狗負責不讓人碰機器。」—— Warren Bennis
What Is Paperclip?
Paperclip is an open-source orchestration platform for AI agent companies. If tools like Claude Code, OpenClaw, or Codex are individual employees, Paperclip is the company that hires, organizes, and manages them.
Key facts:
- Repo: paperclipai/paperclip
- Stack: Node.js + Express 5 + React + PostgreSQL + Drizzle ORM
- License: MIT
- Stars: ~26,900 (in two weeks)
- Status: v0.3.x, pre-1.0, very actively developed
- Created: March 2, 2026
What Paperclip is NOT:
- Not a chatbot
- Not an agent framework (like CrewAI or LangGraph)
- Not a workflow builder (like n8n or Zapier)
- Not a prompt manager
It is a control plane — a single dashboard where you define company goals, build org charts, assign agents to roles, set budgets, schedule heartbeats, and maintain board-level governance over everything.
Paperclip 是什麼?
Paperclip 是一個開源的 AI Agent 公司編排平台。如果 Claude Code、OpenClaw、Codex 是個別的員工,Paperclip 就是僱用、組織、管理他們的公司。
重點資訊:
- 儲存庫: paperclipai/paperclip
- 技術棧: Node.js + Express 5 + React + PostgreSQL + Drizzle ORM
- 授權: MIT
- 星數: 約 26,900(兩週內)
- 狀態: v0.3.x,pre-1.0,非常積極開發中
- 建立日期: 2026 年 3 月 2 日
Paperclip 不是什麼:
- 不是聊天機器人
- 不是 agent 框架(像 CrewAI 或 LangGraph)
- 不是工作流程建構器(像 n8n 或 Zapier)
- 不是 prompt 管理器
它是一個 control plane——一個單一儀表板,你在這裡定義公司目標、建立組織架構、指派 agent 到角色、設定預算、排程 heartbeat,並對一切維持董事會層級的治理。
Architecture Overview
┌─────────────────────────────────────────────────┐
│ Paperclip │
│ │
│ ┌──────────┐ ┌──────────┐ ┌───────────────┐ │
│ │ React UI │ │ Express │ │ PostgreSQL │ │
│ │ (Dashboard)│ │ API Server│ │ (embedded or │ │
│ │ │ │ │ │ external) │ │
│ └────┬─────┘ └────┬─────┘ └───────────────┘ │
│ │ │ │
│ └──────┬───────┘ │
│ │ │
│ ┌───────────▼────────────┐ │
│ │ Agent Adapters │ │
│ │ ┌──────┐ ┌─────────┐ │ │
│ │ │Claude│ │OpenClaw │ │ │
│ │ │Code │ │Gateway │ │ │
│ │ ├──────┤ ├─────────┤ │ │
│ │ │Codex │ │Cursor │ │ │
│ │ ├──────┤ ├─────────┤ │ │
│ │ │Gemini│ │Process │ │ │
│ │ └──────┘ └─────────┘ │ │
│ └────────────────────────┘ │
└─────────────────────────────────────────────────┘Each agent type has an adapter that implements three methods: invoke(), status(), and cancel(). This means Paperclip doesn't care what the agent is — it only cares that the adapter can start it, check on it, and stop it.
Supported adapters:
| Adapter | Type | Description |
|---|---|---|
claude-local | CLI | Claude Code running locally |
codex-local | CLI | OpenAI Codex running locally |
cursor-local | CLI | Cursor running locally |
gemini-local | CLI | Gemini CLI running locally |
openclaw-gateway | WebSocket | OpenClaw via gateway connection |
opencode-local | CLI | OpenCode running locally |
pi-local | CLI | Pi running locally |
process | Shell | Any shell command |
http | Webhook | Any HTTP endpoint |
架構概覽
┌─────────────────────────────────────────────────┐
│ Paperclip │
│ │
│ ┌──────────┐ ┌──────────┐ ┌───────────────┐ │
│ │ React UI │ │ Express │ │ PostgreSQL │ │
│ │(儀表板) │ │ API 伺服器│ │ (內嵌或外部)│ │
│ │ │ │ │ │ │ │
│ └────┬─────┘ └────┬─────┘ └───────────────┘ │
│ │ │ │
│ └──────┬───────┘ │
│ │ │
│ ┌───────────▼────────────┐ │
│ │ Agent Adapters │ │
│ │ ┌──────┐ ┌─────────┐ │ │
│ │ │Claude│ │OpenClaw │ │ │
│ │ │Code │ │Gateway │ │ │
│ │ ├──────┤ ├─────────┤ │ │
│ │ │Codex │ │Cursor │ │ │
│ │ ├──────┤ ├─────────┤ │ │
│ │ │Gemini│ │Process │ │ │
│ │ └──────┘ └─────────┘ │ │
│ └────────────────────────┘ │
└─────────────────────────────────────────────────┘每種 agent 類型有一個 adapter,實作三個方法:invoke()、status() 和 cancel()。這意味著 Paperclip 不在乎 agent 是什麼——它只在乎 adapter 能啟動它、檢查它、停止它。
支援的 adapter:
| Adapter | 類型 | 說明 |
|---|---|---|
claude-local | CLI | 本地執行的 Claude Code |
codex-local | CLI | 本地執行的 OpenAI Codex |
cursor-local | CLI | 本地執行的 Cursor |
gemini-local | CLI | 本地執行的 Gemini CLI |
openclaw-gateway | WebSocket | 透過 gateway 連線的 OpenClaw |
opencode-local | CLI | 本地執行的 OpenCode |
pi-local | CLI | 本地執行的 Pi |
process | Shell | 任意 shell 指令 |
http | Webhook | 任意 HTTP 端點 |
Deployment: Three Ways
Option A: One-Line Quick Start (Local)
npx paperclipai onboard --yesThis handles everything: clones the repo, installs dependencies, provisions an embedded PostgreSQL, and starts the server. Open http://localhost:3100.
Requirements: Node.js 20+, npm/npx available.
Option B: Manual Clone (Local Development)
git clone https://github.com/paperclipai/paperclip.git
cd paperclip
pnpm install
pnpm devRequirements: Node.js 20+, pnpm 9.15+.
The embedded PostgreSQL is auto-provisioned — no external database needed for local development. Data lives under ~/.paperclip/instances/default/.
Option C: Docker
# Build
docker build -t paperclip-local .
# Run
docker run --name paperclip \
-p 3100:3100 \
-e HOST=0.0.0.0 \
-e PAPERCLIP_HOME=/paperclip \
-v "$(pwd)/data/docker-paperclip:/paperclip" \
paperclip-localOr use Docker Compose:
docker compose -f docker-compose.quickstart.yml upThe volume mount (-v) ensures your data persists across container restarts.
部署:三種方式
方案 A:一行指令快速啟動(本地)
npx paperclipai onboard --yes這會處理所有事:clone 儲存庫、安裝相依套件、佈建內嵌 PostgreSQL,然後啟動伺服器。開啟 http://localhost:3100。
需求: Node.js 20+,npm/npx 可用。
方案 B:手動 Clone(本地開發)
git clone https://github.com/paperclipai/paperclip.git
cd paperclip
pnpm install
pnpm dev需求: Node.js 20+,pnpm 9.15+。
內嵌 PostgreSQL 會自動佈建——本地開發不需要外部資料庫。資料存放在 ~/.paperclip/instances/default/ 底下。
方案 C:Docker
# 建構
docker build -t paperclip-local .
# 執行
docker run --name paperclip \
-p 3100:3100 \
-e HOST=0.0.0.0 \
-e PAPERCLIP_HOME=/paperclip \
-v "$(pwd)/data/docker-paperclip:/paperclip" \
paperclip-local或使用 Docker Compose:
docker compose -f docker-compose.quickstart.yml upVolume 掛載(-v)確保你的資料在容器重啟後仍然存在。
Deployment Modes: Security Levels
Paperclip has three deployment modes that control authentication and network exposure:
| Mode | Auth | Binding | Use Case |
|---|---|---|---|
local_trusted | No login required | Loopback only (127.0.0.1) | Solo operator, local machine |
authenticated + private | Session-based login | LAN / VPN / Tailscale | Private network access |
authenticated + public | Session-based login | Internet-facing | Cloud deployment |
Configure via environment variables:
# Local trusted (default)
PAPERCLIP_DEPLOYMENT_MODE=local_trusted
# Authenticated on private network
PAPERCLIP_DEPLOYMENT_MODE=authenticated
PAPERCLIP_DEPLOYMENT_EXPOSURE=private
# Authenticated on public internet
PAPERCLIP_DEPLOYMENT_MODE=authenticated
PAPERCLIP_DEPLOYMENT_EXPOSURE=public
PAPERCLIP_PUBLIC_URL=https://your-domain.comRecommendation for solo operators: Start with local_trusted. When you need remote access from your phone or another machine, switch to authenticated + private behind Tailscale. Only use public when absolutely necessary.
部署模式:安全等級
Paperclip 有三種部署模式,控制認證和網路曝露:
| 模式 | 認證 | 綁定 | 使用場景 |
|---|---|---|---|
local_trusted | 不需要登入 | 僅 loopback(127.0.0.1) | 單人操作,本地機器 |
authenticated + private | Session 登入 | LAN / VPN / Tailscale | 私有網路存取 |
authenticated + public | Session 登入 | 面向網際網路 | 雲端部署 |
透過環境變數設定:
# 本地信任(預設)
PAPERCLIP_DEPLOYMENT_MODE=local_trusted
# 私有網路認證
PAPERCLIP_DEPLOYMENT_MODE=authenticated
PAPERCLIP_DEPLOYMENT_EXPOSURE=private
# 公開網路認證
PAPERCLIP_DEPLOYMENT_MODE=authenticated
PAPERCLIP_DEPLOYMENT_EXPOSURE=public
PAPERCLIP_PUBLIC_URL=https://your-domain.com給單人操作者的建議: 從 local_trusted 開始。當你需要從手機或其他機器遠端存取時,切換到 authenticated + private 並搭配 Tailscale。只在絕對必要時才使用 public。
Production Deployment
For production environments, swap embedded PostgreSQL for an external instance and configure persistent storage:
1. External PostgreSQL
DATABASE_URL=postgresql://user:password@your-db-host:5432/paperclip2. S3 Storage (Optional)
For file storage at scale, configure AWS S3:
pnpm paperclipai configure --section storageWithout S3, files are stored locally under ~/.paperclip/.
3. Health Check
pnpm paperclipai doctorThis diagnoses configuration issues — database connectivity, missing API keys, adapter availability.
4. Key Environment Variables
| Variable | Default | Description |
|---|---|---|
DATABASE_URL | (embedded) | PostgreSQL connection string |
PORT | 3100 | Server port |
HOST | 127.0.0.1 | Bind address |
PAPERCLIP_HOME | ~/.paperclip | Data directory |
PAPERCLIP_INSTANCE_ID | default | Instance identifier |
SERVE_UI | true | Whether to serve the React UI |
ANTHROPIC_API_KEY | — | For Claude Code adapter |
OPENAI_API_KEY | — | For Codex adapter |
生產環境部署
在生產環境中,將內嵌 PostgreSQL 替換為外部實例並設定持久化儲存:
1. 外部 PostgreSQL
DATABASE_URL=postgresql://user:password@your-db-host:5432/paperclip2. S3 儲存(選用)
大規模檔案儲存可設定 AWS S3:
pnpm paperclipai configure --section storage沒有 S3 的話,檔案會儲存在 ~/.paperclip/ 本地目錄下。
3. 健康檢查
pnpm paperclipai doctor這會診斷設定問題——資料庫連線、缺少的 API key、adapter 可用性。
4. 關鍵環境變數
| 變數 | 預設值 | 說明 |
|---|---|---|
DATABASE_URL | (內嵌) | PostgreSQL 連線字串 |
PORT | 3100 | 伺服器連接埠 |
HOST | 127.0.0.1 | 綁定位址 |
PAPERCLIP_HOME | ~/.paperclip | 資料目錄 |
PAPERCLIP_INSTANCE_ID | default | 實例識別碼 |
SERVE_UI | true | 是否提供 React UI |
ANTHROPIC_API_KEY | — | 供 Claude Code adapter 使用 |
OPENAI_API_KEY | — | 供 Codex adapter 使用 |
How to Use Paperclip: Step by Step
Step 1: Create a Company
After deployment, open the dashboard and create a company. Define the core goal:
- "Build an AI note-taking app to $1M MRR"
- "Maintain and ship features for our open-source project"
- "Run content marketing for a SaaS product"
The goal drives everything — task hierarchies, agent priorities, and budget allocation all trace back to it.
Step 2: Build the Org Chart
Hire agents into roles. The hierarchy typically looks like:
Board (You — the human operator)
└── CEO (strategic agent)
├── CTO
│ ├── Backend Engineer
│ ├── Frontend Engineer
│ └── DevOps Engineer
├── Head of Marketing
│ ├── Content Writer
│ └── Social Media Manager
└── Head of Product
└── Product AnalystEach role maps to an agent with a specific adapter. For example:
- CTO → Claude Code (
claude-localadapter) - Backend Engineer → Codex (
codex-localadapter) - Content Writer → OpenClaw (
openclaw-gatewayadapter)
Step 3: Configure Budgets
Set monthly token or dollar budgets per agent:
- Soft alert at 80% — you get notified
- Hard ceiling at 100% — agent auto-pauses
This prevents cost spirals when agents get stuck in loops or generate excessive API calls.
Step 4: Set Heartbeat Schedules
Agents don't run constantly. They wake on a schedule (the "heartbeat"), check their assigned tasks, work, and report back. Configure frequency per agent:
- CTO: Every 2 hours — reviews architecture decisions, assigns tasks
- Engineers: Every 30 minutes — pick up issues, write code, submit PRs
- Content Writer: Every 4 hours — draft articles, schedule posts
Step 5: Approve and Monitor
The Board (you) approves:
- Strategy changes — before the CEO pivots direction
- New hires — before adding agents to the org chart
- Budget increases — before raising spending limits
The dashboard shows real-time status: which agents are active, what tasks they're working on, current spend, and conversation logs.
Task Hierarchy
All work follows a structured hierarchy that traces back to the company goal:
Company Goal
└── Initiative (strategic)
└── Project
└── Milestone
└── Issue (assignable to agents)
└── Sub-issueAgents check out issues atomically — no two agents work on the same issue simultaneously.
如何使用 Paperclip:逐步指南
第一步:建立公司
部署完成後,開啟儀表板並建立一家公司。定義核心目標:
- 「建構一個 AI 筆記應用程式,達到 $1M MRR」
- 「維護我們的開源專案並交付新功能」
- 「為 SaaS 產品執行內容行銷」
目標驅動一切——任務層級、agent 優先順序、預算分配都追溯回它。
第二步:建立組織架構
僱用 agent 到各個角色。層級通常看起來像:
Board(你——人類操作者)
└── CEO(策略 agent)
├── CTO
│ ├── Backend Engineer
│ ├── Frontend Engineer
│ └── DevOps Engineer
├── Head of Marketing
│ ├── Content Writer
│ └── Social Media Manager
└── Head of Product
└── Product Analyst每個角色對應一個帶有特定 adapter 的 agent。例如:
- CTO → Claude Code(
claude-localadapter) - Backend Engineer → Codex(
codex-localadapter) - Content Writer → OpenClaw(
openclaw-gatewayadapter)
第三步:設定預算
為每個 agent 設定每月 token 或美元預算:
- 80% 軟警報 —— 你會收到通知
- 100% 硬上限 —— agent 自動暫停
這能防止 agent 卡在迴圈或產生過多 API 呼叫時的成本失控。
第四步:設定 Heartbeat 排程
Agent 不會持續運行。它們按排程(「heartbeat」)醒來,檢查被分配的任務,工作,然後回報。為每個 agent 設定頻率:
- CTO: 每 2 小時——審查架構決策、分配任務
- 工程師: 每 30 分鐘——領取 issue、寫程式碼、提交 PR
- Content Writer: 每 4 小時——撰寫文章、排程發佈
第五步:核准與監控
Board(你)核准:
- 策略變更 —— 在 CEO 改變方向之前
- 新僱用 —— 在新增 agent 到組織架構之前
- 預算增加 —— 在提高花費上限之前
儀表板顯示即時狀態:哪些 agent 正在活躍、正在處理哪些任務、目前花費、以及對話紀錄。
任務層級
所有工作遵循結構化的層級,追溯回公司目標:
Company Goal
└── Initiative(策略級)
└── Project
└── Milestone
└── Issue(可分配給 agent)
└── Sub-issueAgent 以原子方式 checkout issue——沒有兩個 agent 會同時處理同一個 issue。
Real-World Usage Patterns
Pattern 1: Solo Developer Scaling
You're a solo developer who needs to ship faster. Deploy Paperclip locally with:
- 1 CTO agent (Claude Code) — architecture decisions, code review
- 2 Engineer agents (Claude Code + Codex) — parallel feature development in isolated worktrees
- 1 DevOps agent (Process adapter) — CI/CD monitoring, deployment scripts
Total cost: your existing API keys. No additional infrastructure.
Pattern 2: Content Operation
Running a blog, newsletter, or social media presence:
- 1 Content Strategist (OpenClaw) — plans editorial calendar
- 2 Writers (OpenClaw) — draft articles
- 1 Editor (Claude Code) — review and polish drafts
Heartbeat schedule: Strategist weekly, Writers daily, Editor triggers on new drafts.
Pattern 3: Open-Source Project Maintenance
Maintaining a popular open-source project with limited time:
- 1 Triage agent — labels and prioritizes incoming issues
- 1 PR Review agent (Claude Code) — reviews pull requests, suggests changes
- 1 Docs agent — keeps documentation in sync with code changes
- 1 Release agent (Process adapter) — manages changelog and release workflow
Multi-Company Support
One Paperclip deployment can host multiple companies with complete data isolation. This is useful if you run several projects or businesses — each gets its own org chart, budget, and agent pool.
實際使用場景
場景 1:單人開發者擴展
你是一個需要更快交付的單人開發者。在本地部署 Paperclip:
- 1 個 CTO agent(Claude Code)—— 架構決策、code review
- 2 個 Engineer agent(Claude Code + Codex)—— 在隔離的 worktree 中平行開發功能
- 1 個 DevOps agent(Process adapter)—— CI/CD 監控、部署腳本
總成本:你現有的 API key。不需要額外基礎設施。
場景 2:內容營運
經營部落格、電子報、或社群媒體:
- 1 個 Content Strategist(OpenClaw)—— 規劃編輯日曆
- 2 個 Writer(OpenClaw)—— 撰寫文章草稿
- 1 個 Editor(Claude Code)—— 審閱並潤飾草稿
Heartbeat 排程:Strategist 每週、Writer 每天、Editor 在有新草稿時觸發。
場景 3:開源專案維護
用有限的時間維護一個熱門開源專案:
- 1 個 Triage agent —— 標記和排序收到的 issue
- 1 個 PR Review agent(Claude Code)—— 審查 pull request、建議修改
- 1 個 Docs agent —— 讓文件與程式碼變更保持同步
- 1 個 Release agent(Process adapter)—— 管理 changelog 和發佈流程
多公司支援
一個 Paperclip 部署可以託管多家公司,且有完全的資料隔離。如果你經營多個專案或事業,每個都有自己的組織架構、預算和 agent 池,這非常有用。
Security Considerations
1. Start With local_trusted, Upgrade When Needed
The default mode binds to loopback only and requires no authentication. This is safe for a single-operator laptop. Don't jump to public unless you have a clear reason.
2. Board Claim Flow
When migrating from local_trusted to authenticated, Paperclip emits a one-time high-entropy claim URL in the server logs. Use it to claim board ownership. This URL is single-use — if someone else claims it, they become the board operator.
Action: Watch the server logs carefully during migration. Claim immediately.
3. Agent API Keys
Each agent adapter may need its own API key (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.). These are stored in Paperclip's configuration.
- Set them via environment variables, not hardcoded in config files
- Never commit
.envfiles to version control - Use separate API keys for Paperclip agents vs. your personal usage — this makes cost attribution easier
4. Budget Enforcement Is Your Safety Net
Without budgets, a misconfigured agent can burn through hundreds of dollars in API calls. Always set hard ceilings:
Budget: $50/month per agent
Soft alert: 80% ($40)
Hard ceiling: 100% ($50, auto-pause)Start conservative. Increase after you observe actual usage patterns.
5. Audit Trail
Every conversation, decision, and tool call is recorded in an immutable log. This is critical for:
- Understanding what an agent did and why
- Debugging unexpected behavior
- Cost forensics when a budget alert fires
6. Plugin Trust Model
Paperclip has a plugin system (partially implemented). Currently, plugin UI runs as same-origin JavaScript — meaning plugins have the same access as the main UI. Only install plugins you trust. There is no sandbox yet.
7. No Multi-User RBAC Yet
V0.3.x is single-operator. There is no role-based access control for multiple human users. If you need multi-user access, wait for the post-V1 roadmap or implement network-level access control (VPN, Tailscale ACLs).
安全考量
1. 從 local_trusted 開始,需要時再升級
預設模式僅綁定 loopback 且不需要認證。對單人操作的筆電來說這是安全的。除非有明確理由,否則不要跳到 public。
2. Board 認領流程
從 local_trusted 遷移到 authenticated 時,Paperclip 會在伺服器日誌中發出一個一次性高熵認領 URL。用它來認領 board 的所有權。這個 URL 是一次性的——如果別人先認領了,他們就成為 board 操作者。
行動: 遷移期間仔細觀察伺服器日誌。立即認領。
3. Agent API Key
每個 agent adapter 可能需要自己的 API key(ANTHROPIC_API_KEY、OPENAI_API_KEY 等)。這些儲存在 Paperclip 的設定中。
- 透過環境變數設定,不要寫死在設定檔中
- 絕不將
.env檔案提交到版本控制 - 為 Paperclip agent 使用和你個人使用不同的 API key——這讓成本歸因更容易
4. 預算控制是你的安全網
沒有預算的話,一個設定錯誤的 agent 可以在 API 呼叫上燒掉幾百美元。永遠設定硬上限:
預算:每個 agent $50/月
軟警報:80%($40)
硬上限:100%($50,自動暫停)保守地開始。在觀察到實際使用模式後再增加。
5. 稽核軌跡
每一次對話、決策、工具呼叫都被記錄在不可變的日誌中。這對以下情況至關重要:
- 理解 agent 做了什麼以及為什麼
- 除錯非預期行為
- 預算警報觸發時的成本鑑識
6. Plugin 信任模型
Paperclip 有一個 plugin 系統(部分實作)。目前 plugin UI 以 same-origin JavaScript 執行——意味著 plugin 和主 UI 有相同的存取權限。只安裝你信任的 plugin。目前還沒有 sandbox。
7. 尚無多使用者 RBAC
V0.3.x 是單人操作者模式。沒有多人類使用者的角色型存取控制。如果你需要多使用者存取,等 post-V1 路線圖或實作網路層級的存取控制(VPN、Tailscale ACL)。
Comparison With Alternatives
| Tool | What It Is | vs. Paperclip |
|---|---|---|
| Claude Code / OpenClaw / Codex | Individual AI agents | Paperclip orchestrates these as employees |
| CrewAI / AutoGen / LangGraph | Agent frameworks | Define how agents work internally; Paperclip defines how a company of agents is governed |
| Asana / Linear / Jira | Task management | Paperclip has built-in task management plus agent-specific features: heartbeats, budgets, atomic checkout |
| n8n / Zapier / Temporal | Workflow automation | Paperclip models companies, not pipelines |
| Kubernetes | Container orchestration | Orchestrates infrastructure; Paperclip orchestrates business logic |
The key distinction: Paperclip treats the company as the fundamental abstraction — org charts, reporting lines, goal hierarchies, cost attribution, and board governance — rather than treating agents as isolated tools or workflow nodes.
與替代方案比較
| 工具 | 是什麼 | vs. Paperclip |
|---|---|---|
| Claude Code / OpenClaw / Codex | 個別 AI agent | Paperclip 將這些編排為員工 |
| CrewAI / AutoGen / LangGraph | Agent 框架 | 定義 agent 內部如何運作;Paperclip 定義 agent 公司如何被治理 |
| Asana / Linear / Jira | 任務管理 | Paperclip 有內建任務管理,外加 agent 特有功能:heartbeat、預算、原子 checkout |
| n8n / Zapier / Temporal | 工作流程自動化 | Paperclip 建模公司,不是管線 |
| Kubernetes | 容器編排 | 編排基礎設施;Paperclip 編排商業邏輯 |
關鍵區別:Paperclip 將公司作為基本抽象——組織架構、報告線、目標層級、成本歸因和董事會治理——而非將 agent 視為孤立的工具或工作流程節點。
Known Limitations
As of v0.3.1 (March 2026):
- Pre-1.0 stability — expect breaking changes between releases
- No multi-user RBAC — single operator only; multi-member boards are post-V1
- Plugin sandbox missing — plugins run with full same-origin access
- Cloud agent support incomplete — Cursor cloud and e2b are on the roadmap
- Documentation is sparse — "better docs" is explicitly on the roadmap
- Single-node assumption — embedded Postgres and local storage mean horizontal scaling needs manual infrastructure work
- No external ticket system integration yet — Jira, Linear, GitHub Issues sync are planned
- 666 open issues — rapid adoption means many bug reports and feature requests
Recommendation: Pin your version. Update intentionally. Back up ~/.paperclip/ before upgrading.
已知限制
截至 v0.3.1(2026 年 3 月):
- Pre-1.0 穩定性 —— 版本間可能有破壞性變更
- 無多使用者 RBAC —— 僅單人操作者;多成員 board 是 post-V1 目標
- Plugin sandbox 缺失 —— plugin 以完整 same-origin 存取權限執行
- 雲端 agent 支援不完整 —— Cursor cloud 和 e2b 在路線圖中
- 文件稀疏 —— 「更好的文件」明確列在路線圖上
- 單節點假設 —— 內嵌 Postgres 和本地儲存意味著水平擴展需要手動基礎設施工作
- 尚無外部工單系統整合 —— Jira、Linear、GitHub Issues 同步已規劃
- 666 個開放 issue —— 快速採用意味著大量 bug 回報和功能請求
建議: 鎖定版本。有意識地更新。升級前備份 ~/.paperclip/。
Quick Start Checklist
- Install:
npx paperclipai onboard --yes - Open:
http://localhost:3100 - Create a company with a clear goal
- Hire a CEO agent — start with Claude Code (
claude-local) - Set a budget — $50/month hard ceiling to start
- Configure heartbeat — every 2 hours for strategic agents, 30 minutes for execution agents
- Add one engineer agent — observe before scaling
- Monitor the dashboard — watch costs, task progress, conversation logs
- Scale gradually — add agents only when the current team is saturated
Start small. One company, two agents, tight budgets. Expand after you trust the system.
快速啟動檢查清單
- 安裝:
npx paperclipai onboard --yes - 開啟:
http://localhost:3100 - 建立公司,訂定清晰的目標
- 僱用 CEO agent —— 從 Claude Code(
claude-local)開始 - 設定預算 —— 起初設 $50/月硬上限
- 設定 heartbeat —— 策略 agent 每 2 小時,執行 agent 每 30 分鐘
- 增加一個 engineer agent —— 觀察後再擴展
- 監控儀表板 —— 觀察成本、任務進度、對話紀錄
- 逐步擴展 —— 只在現有團隊飽和時才增加 agent
從小開始。一家公司、兩個 agent、緊縮預算。信任系統之後再擴展。