30 seconds to your AI company 🦞
# Option A: npm global install
npm install -g @jackclaw/cli
jackclaw demo
# Option B: git clone
git clone https://github.com/DevJackKong/JackClawOS.git
cd JackClawOS && npm install && npm run build && npx jackclaw demo→ Hub ready on :3100 · Node ready on :19000 · Dashboard at :3100 · Quick Start
JackClaw is an open-source multi-agent collaboration framework that lets you run a fleet of AI agents as if they were employees at a company. Each agent has its own encrypted identity, persistent 4-layer memory, permission level, and communication channel. A central Hub orchestrates task routing, enforces human-in-the-loop approval for high-risk actions, and aggregates daily reports from every agent.
Built on TypeScript with end-to-end RSA-4096 + AES-256 encryption, JackClaw is designed to plug into your existing workflow — whether you're running Claude Code, Codex, or Cursor. The entire system starts with a single command: jackclaw demo.
Key design principles:
- You are always in control — high-risk actions (payments, deploys, deletes) require explicit CEO approval
- Privacy by default — each agent's memory is private; the Hub only sees encrypted ciphertext
- Production-grade security — tamper-proof audit logs, per-agent trust scoring, rate limiting
- Zero lock-in — MIT licensed, self-hosted, works with any LLM provider (16 supported)
| Feature | JackClaw | Single-Agent Tools |
|---|---|---|
| Multi-agent collaboration | ✅ Built-in | ❌ |
| Agent-to-agent messaging | ✅ ClawChat (WS + REST) | ❌ |
| Daily reports & summaries | ✅ Auto-aggregated | ❌ |
| Trust & reputation system | ✅ Per-agent scoring | ❌ |
| CEO approval for high-risk ops | ✅ Human-in-loop | ❌ |
| Web dashboard | ✅ Real-time | ❌ |
| One-command demo | ✅ jackclaw demo |
❌ |
| End-to-end encryption | ✅ RSA-4096 + AES-256 | Varies |
| OpenClaw compatible | ✅ Plugin | N/A |
# See it in action
npx jackclaw demoJackClaw 是一个让多个 AI Agent 像公司员工一样协作的框架。
你是 CEO，AI 是你的团队。每个 AI 有自己的记忆、技能和权限。你下指令，它们干活、汇报、互相配合。高风险操作（花钱、发布、删除）必须你亲自批准。
现在的 AI 工具有一个问题：它们是孤立的。
- ChatGPT 不知道你的 Cursor 在写什么代码
- Claude 不知道你的另一个 Claude 昨天做了什么
- 你每天花大量时间在不同 AI 之间复制粘贴上下文
想象一下，如果你的 AI 们能像真人团队一样：
- 🧑💻 工程师 AI 写好代码，自动通知测试 AI 去验证
- 📊 分析师 AI 发现数据异常，直接告诉运维 AI 去排查
- 💰 财务 AI 需要付款，必须等你点「批准」才能执行
- 📝 所有 AI 每天给你发工作日报
这就是 JackClaw 做的事。
你（CEO）
│
▼
┌─────────────────────────────────┐
│ Hub（总部） │
│ │
│ • 接收你的指令 │
│ • 把任务分给对的 AI │
│ • 汇总所有 AI 的日报 │
│ • 拦截高风险操作等你批准 │
│ • 所有通信加密，没人能偷看 │
└──────┬──────────┬──────────┬────┘
│ │ │
┌───▼───┐ ┌──▼───┐ ┌──▼───┐
│ AI-1 │ │ AI-2 │ │ AI-3 │
│ 后端 │ │ 前端 │ │ 运维 │
│ 工程师│ │ 设计 │ │ 部署 │
└───────┘ └──────┘ └──────┘
各自有私有记忆，互相教技能
三个角色：
| 角色 | 是谁 | 干什么 |
|---|---|---|
| 你 | CEO / 老板 | 下指令，审批高风险操作 |
| Hub | 总部 / 调度中心 | 分活、汇报、安全管控 |
| Node | AI 员工 | 干活、学习、每天汇报 |
每个 AI 有四层记忆：
| 层 | 比喻 | 作用 |
|---|---|---|
| L1 | 工作记忆 | 当前在干什么（毫秒级） |
| L2 | 笔记本 | 昨天学到了什么（本地存储） |
| L3 | 公司知识库 | 团队共享的知识（需授权） |
| L4 | 云端备份 | 同步到总部（你控制） |
AI-1 教会了 AI-2 一个技能？AI-2 会永远记住。 不像现在，关掉窗口就全忘了。
JackClaw 有四个权限等级：
| 等级 | 意思 | 例子 |
|---|---|---|
| L0 受监督 | AI 只能看，不能动 | 新来的实习生 |
| L1 辅助 | 能查询，不能改 | 刚入职的员工 |
| L2 标准 | 能读能写，但不能做大事 | 正式员工 |
| L3 自主 | 几乎什么都能做 | 信任的高管 |
但无论什么等级，花钱、删除、发布这些操作，永远需要你批准。
不是复制粘贴，是真正的协作：
- 任务传递：Hub 把任务分给最合适的 AI
- 即时通讯：AI 之间有加密聊天频道（ClawChat）
- 技能转移：擅长写 Python 的 AI 可以教另一个 AI
- 协作会话：两个 AI 可以共享一段记忆来合作完成任务
- 所有通信 RSA-4096 + AES-256 加密（和银行一个级别）
- Hub 只转发密文，自己也看不到内容
- 每个 AI 的记忆是私有的，其他 AI 和 Hub 都读不了
- 所有高风险操作有不可篡改的审计日志
# 30 秒创建一个 AI 团队
npm create jackclaw@latest my-team
cd my-team
# 启动总部
jackclaw start --role hub
# 加一个 AI 员工
jackclaw start --role node --name "工程师小明"
# 给它派活
jackclaw chat --to @工程师小明 --text "帮我写个登录页面"支持接入 Claude Code、Codex、Cursor 等主流 AI 工具。
| 你是谁 | 用 JackClaw 做什么 |
|---|---|
| 独立开发者 | 一个人管理多个 AI 助手，像有了一个小团队 |
| 创业者 | 用 AI 团队代替早期招人，降低成本 |
| 技术团队 | 让 AI 处理重复工作，人类专注创造性任务 |
| AI 爱好者 | 探索多 Agent 协作的前沿玩法 |
| ChatGPT | AutoGPT | JackClaw | |
|---|---|---|---|
| 多 Agent 协作 | ❌ | 有限 | ✅ 完整 |
| 持久记忆 | 有限 | ❌ | ✅ 四层 |
| 人类审批 | ❌ | ❌ | ✅ 内置 |
| 加密通信 | ❌ | ❌ | ✅ 军事级 |
| IDE 集成 | ❌ | ❌ | ✅ Claude/Codex/Cursor |
| 开源免费 | ❌ | ✅ | ✅ MIT |
npm install -g @jackclaw/cli
jackclaw demo📦 GitHub：github.com/DevJackKong/JackClawOS 📄 协议：MIT（免费，随便用）
JackClaw is a framework that lets multiple AI agents collaborate like a real company.
You're the CEO. AI agents are your team. Each has its own memory, skills, and permissions. You give orders, they work, report back, and help each other. High-risk actions (spending money, deploying, deleting) require your explicit approval.
Today's AI tools have a problem: they're isolated.
- ChatGPT doesn't know what your Cursor is coding
- Claude doesn't know what your other Claude did yesterday
- You spend hours copy-pasting context between AI tools
Imagine if your AI agents could work like a real team:
- 🧑💻 Engineer AI finishes code → automatically notifies Tester AI to verify
- 📊 Analyst AI spots a data anomaly → tells DevOps AI to investigate
- 💰 Finance AI needs to make a payment → waits for your "Approve" before executing
- 📝 Every AI sends you a daily work report
That's what JackClaw does.
You (CEO)
│
▼
┌─────────────────────────────────┐
│ Hub (HQ) │
│ │
│ • Receives your commands │
│ • Routes tasks to the right AI │
│ • Aggregates daily reports │
│ • Blocks high-risk actions │
│ • All comms encrypted │
└──────┬──────────┬──────────┬────┘
│ │ │
┌───▼───┐ ┌──▼───┐ ┌──▼───┐
│ AI-1 │ │ AI-2 │ │ AI-3 │
│Backend│ │Front │ │DevOps│
│ Dev │ │ End │ │ Ops │
└───────┘ └──────┘ └──────┘
Each has private memory, can teach each other skills
Three roles:
| Role | Who | What They Do |
|---|---|---|
| You | CEO / Owner | Give orders, approve high-risk actions |
| Hub | HQ / Dispatcher | Route tasks, collect reports, enforce security |
| Node | AI Employee | Do the work, learn, report daily |
Each AI has four memory layers:
| Layer | Analogy | Purpose |
|---|---|---|
| L1 | Working memory | What I'm doing right now (<5ms) |
| L2 | Notebook | What I learned yesterday (local) |
| L3 | Company wiki | Shared team knowledge (opt-in) |
| L4 | Cloud backup | Synced to HQ (you control) |
AI-1 teaches AI-2 a skill? AI-2 remembers forever. Unlike now, where closing the tab means starting over.
Four permission levels:
| Level | Meaning | Like... |
|---|---|---|
| L0 Supervised | Read-only | New intern |
| L1 Assisted | Can query, can't change | Junior hire |
| L2 Standard | Read + write, no big moves | Regular employee |
| L3 Autonomous | Almost everything | Trusted executive |
But regardless of level, spending money, deleting things, and deploying always need your approval.
Not copy-paste. Real collaboration:
- Task routing: Hub assigns tasks to the best-fit AI
- Instant messaging: Encrypted chat between AIs (ClawChat)
- Skill transfer: A Python expert AI can teach another AI
- Collab sessions: Two AIs share memory to work on a task together
- All communication encrypted with RSA-4096 + AES-256 (bank-level)
- Hub only relays ciphertext — it can't read the content
- Each AI's memory is private — other AIs and the Hub can't access it
- Tamper-proof audit logs for all high-risk actions
# Create an AI team in 30 seconds
npm create jackclaw@latest my-team
cd my-team
# Start HQ
jackclaw start --role hub
# Add an AI employee
jackclaw start --role node --name "engineer-alex"
# Assign work
jackclaw chat --to @engineer-alex --text "Build me a login page"Works with Claude Code, Codex, Cursor, and more.
| You Are... | Use JackClaw To... |
|---|---|
| Solo developer | Manage multiple AI assistants like having a small team |
| Startup founder | Replace early hires with an AI team, cut costs |
| Tech team | Let AI handle repetitive work, humans focus on creative tasks |
| AI enthusiast | Explore cutting-edge multi-agent collaboration |
| ChatGPT | AutoGPT | JackClaw | |
|---|---|---|---|
| Multi-agent collab | ❌ | Limited | ✅ Full |
| Persistent memory | Limited | ❌ | ✅ 4-layer |
| Human approval | ❌ | ❌ | ✅ Built-in |
| Encrypted comms | ❌ | ❌ | ✅ Military-grade |
| IDE integration | ❌ | ❌ | ✅ Claude/Codex/Cursor |
| Open source & free | ❌ | ✅ | ✅ MIT |
npm install -g @jackclaw/cli
jackclaw demo📦 GitHub: github.com/DevJackKong/JackClawOS 📄 License: MIT (free, use however you want)
Complete list of jackclaw commands:
| Command | Description |
|---|---|
jackclaw demo |
Run a live demo: Hub + 3 AI employees collaborate |
jackclaw start |
Start JackClaw Hub and/or Node services |
jackclaw stop |
Stop the JackClaw daemon |
jackclaw status |
Show node status and Hub connection |
jackclaw init |
Initialize node identity and create config |
jackclaw invite <hub-url> |
Register this node with a Hub |
jackclaw chat |
Open interactive ClawChat session via Hub WebSocket |
jackclaw ask <prompt> |
Send a prompt to any LLM via Hub |
jackclaw task run <prompt> |
Submit a task to a node for LLM execution |
jackclaw task status <id> |
Check the status of a submitted task |
jackclaw send <handle> <msg> |
Send a direct message to @handle |
jackclaw inbox |
View your incoming messages |
jackclaw mention <handle> |
Send a collaboration invitation to another agent |
jackclaw sessions list |
List your active collaboration sessions |
jackclaw sessions respond <id> <decision> |
Accept / decline a session invite |
jackclaw sessions end <id> |
End a collaboration session |
jackclaw identity register <handle> |
Register your @handle on the Hub |
jackclaw identity lookup <handle> |
Look up another agent by @handle |
jackclaw identity who |
Show agents on this Hub |
jackclaw report |
Send a work report to Hub |
jackclaw nodes |
List all nodes (Hub role only) |
jackclaw logs [nodeId] |
View node health and activity via Watchdog |
jackclaw providers |
List available LLM providers across all nodes |
jackclaw model list |
List all available models (local + cloud) |
jackclaw model set <model> |
Set default model (e.g. openai/gpt-4o) |
jackclaw model scan |
Scan for locally available models (Ollama + MLX) |
jackclaw model set-key <provider> <key> |
Configure an API key for a cloud provider |
jackclaw schedule <agent> <time> |
Negotiate a meeting time with another agent |
jackclaw remind <args> |
Create or cancel a reminder |
jackclaw reminders |
View your reminder list |
jackclaw secretary status |
Show AI secretary mode and stats |
jackclaw secretary mode <mode> |
Set secretary mode (auto-reply / summary / off) |
jackclaw translate on|off |
Enable / disable auto-translation of messages |
jackclaw filter status |
Show today's message filter statistics |
jackclaw filter whitelist add <handle> |
Always allow messages from a handle |
jackclaw filter blacklist add <handle> |
Always block messages from a handle |
jackclaw social send <handle> <msg> |
Send a social message through your agent |
jackclaw moltbook connect <api_key> |
Connect a Moltbook social account |
jackclaw config [key] [value] |
View or modify configuration |
jackclaw hub-status |
Show Hub connectivity and online agents |
JackClaw is a monorepo with 14 packages:
| Package | npm | Description |
|---|---|---|
@jackclaw/cli |
|
CLI — jackclaw start / demo / chat |
@jackclaw/hub |
 |
Central orchestrator — REST + WebSocket |
@jackclaw/node |
 |
AI agent worker — register, report, execute |
@jackclaw/protocol |
 |
RSA-4096 + AES-256 encrypted messaging |
@jackclaw/llm-gateway |
 |
Multi-model gateway — 16 providers |
@jackclaw/memory |
 |
4-layer agent memory + semantic search |
@jackclaw/sdk |
 |
Plugin / Node development SDK |
@jackclaw/harness |
 |
ACP harness adapter layer |
@jackclaw/tunnel |
 |
Cloudflared + self-hosted HTTPS tunnel |
@jackclaw/openclaw-plugin |
 |
OpenClaw integration plugin |
@jackclaw/create |
 |
npm create jackclaw@latest scaffolding |
@jackclaw/dashboard |
(private) | React real-time web dashboard |
@jackclaw/watchdog |
(private) | Human-only oversight + heartbeat monitor |
@jackclaw/payment-vault |
(private) | CEO-approval payment compliance engine |
┌──────────────────────────────────────────────────────────────┐
│ CEO (You / Human) │
│ npm install -g @jackclaw/cli → jackclaw demo │
└───────────────────────────┬──────────────────────────────────┘
│ JWT Auth (REST + WebSocket)
┌───────────────────────────▼──────────────────────────────────┐
│ @jackclaw/hub :3100 │
│ │
│ Routes tasks │ Aggregates reports │ Human-review queue │
│ ClawChat WS │ Trust graph │ Dashboard UI │
└───┬────────────┬────────────┬─────────┬────────────┬─────────┘
│ │ │ │ │
│ @jackclaw/protocol (RSA-4096 + AES-256 per message)
│ │ │ │ │
┌───▼───┐ ┌────▼───┐ ┌─────▼──┐ ┌──▼─────┐ ┌──▼─────┐
│ Node1 │ │ Node2 │ │ Node3 │ │ ... │ │ NodeN │
│:19000 │ │:19001 │ │:19002 │ │ │ │ │
└───┬───┘ └────┬───┘ └─────┬──┘ └────────┘ └────────┘
│ │ │
└─ @jackclaw/memory (SQLite, private per node)
└─ @jackclaw/llm-gateway (16 LLM providers)
└─ @jackclaw/harness (ACP adapter)
Side services:
@jackclaw/payment-vault — CEO-approval payment workflow
@jackclaw/watchdog — heartbeat + health metrics
@jackclaw/tunnel — cloudflared public URL (--tunnel flag)
We welcome contributions of all sizes — bug fixes, new features, documentation, and tests.
git clone https://github.com/DevJackKong/JackClawOS.git
cd JackClawOS
npm install # install all workspace dependencies
npm run build # compile all packages (TypeScript → JS)
npm test # run unit tests (286 tests across 14 packages)Start the dev environment:
# Terminal 1 — Hub
node packages/hub/dist/index.js
# Terminal 2 — a sample Node
node packages/node/dist/index.js --name dev-node --hub http://localhost:3100- Language: TypeScript (strict mode, no
anywithout justification) - Style: 2-space indent, single quotes, semicolons — enforced by the existing
tsconfig.json - Tests: add unit tests for all new public APIs; E2E tests for CLI commands
- Security: never log plaintext message content; all inter-node payloads must be encrypted via
@jackclaw/protocol - Commits: use Conventional Commits —
feat:,fix:,docs:,chore:
- Fork the repo and create a branch from
main:git checkout -b feat/my-feature - Make your changes with tests
- Run
npm run build && npm test— all 286 tests must pass - Open a pull request against
mainwith a clear description of what and why - A maintainer will review within 48 hours
- PostgreSQL storage backend for Hub (replacing in-memory state)
- Cross-Hub federation — agents on different Hubs can message each other
- Federated trust graph — reputation scores sync across Hubs
-
jackclaw clustercommand for multi-Hub management
- Plugin marketplace — publish and install community Node plugins
- Agent Memory cloud sync — opt-in L4 memory hosted on JackClaw Cloud
- Visual workflow builder (drag-and-drop task routing in Dashboard)
- Webhook triggers — external events can dispatch tasks to agents
- Full SLA guarantees and stability commitments
- Kubernetes Helm chart for enterprise deployment
- SOC 2 Type II audit trail support
- Official LTS (Long-Term Support) release cadence
JackClaw is released under the MIT License.
MIT License — Copyright (c) 2024 Jack Kong
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
Built by Jack 🦞
One person. Fifty AI agents. That's JackClaw.