swarm-orchestrator

AI Agent cluster orchestration platform - manage, schedule, and coordinate multiple AI agents locally with FastAPI backend and React dashboard

Safety Notice

This listing is from the official public ClawHub registry. Review SKILL.md and referenced scripts before running.

Copy this and send it to your AI assistant to learn

Install skill "swarm-orchestrator" with this command: npx skills add ZhenStaff/swarm-orchestrator

🇨🇳 中文版

Swarm Orchestrator 技能

状态: 🟢 本地优先的 AI 智能体编排平台 类型: 自托管,无需外部依赖 隐私: 100% 本地处理(可选的大模型 API 调用除外)


🔒 安全与信任

本技能的功能

  • 本地运行 - 在您的机器上运行(后端 + 前端)
  • 无遥测 - 不收集数据
  • 无需外部服务器 - 核心功能无需外部依赖
  • 可选 API 密钥 - 仅在使用大模型智能体时需要(OpenAI/Claude)
  • 开源 - 所有代码可审计

本技能不会做的事

  • ❌ 不向外部服务器发送数据(可选的大模型 API 除外)
  • ❌ 不收集分析或遥测数据
  • ❌ 不需要账号注册
  • ❌ 不会未经许可访问您的文件
  • ❌ 不会在您不知情的情况下运行后台进程

数据存储

  • 数据库: SQLite 文件存储在 ./data/swarm.db(仅本地)
  • 日志: 文本文件存储在 ./logs/(仅本地)
  • 缓存: 本地 Redis(仅本地)
  • 无云同步 或远程存储

📋 概述

OpenClaw Swarm Orchestrator 是一个本地优先的平台,用于构建和管理多智能体 AI 系统。可以把它想象成协调多个 AI 智能体协同工作的"控制塔"。

核心功能

  • 智能体注册表 - 注册 LLM、工具、人类和自定义智能体
  • 任务队列 - 基于优先级的任务分发与依赖管理
  • 实时仪表板 - Web UI 监控智能体和任务
  • RESTful API - 完整的 REST API 用于程序化控制
  • 本地存储 - 所有数据保留在您的机器上

架构

┌─────────────────┐
│  Web 仪表板     │  http://localhost:3000
│   (React UI)    │
└────────┬────────┘
         │
┌────────▼────────┐
│  FastAPI 服务器 │  http://localhost:8000
│  (后端 API)     │
└────────┬────────┘
         │
┌────────▼────────┐
│   本地存储      │
│ • SQLite 数据库 │  ./data/swarm.db
│ • Redis 缓存    │  localhost:6379
│ • 日志文件      │  ./logs/*.log
└─────────────────┘

🚀 安装

前置条件检查

安装前,请验证您已安装:

# Python 3.11+
python --version

# Node.js 18+
node --version

# Redis
redis-cli ping  # 应该返回 PONG

# (可选) Docker
docker --version

方法一:使用 Docker(推荐 - 最简单)

这是最安全的方法 - 所有内容都在容器中运行。

# 1. 克隆仓库(先检查代码!)
git clone https://github.com/ZhenRobotics/openclaw-swarm-orchestrator.git
cd openclaw-swarm-orchestrator

# 2. 启动前查看 docker-compose.yml
cat docker-compose.yml

# 3. 启动服务
docker-compose up -d

# 4. 验证
curl http://localhost:8000/health
# 应该返回: {"status": "healthy"}

访问:

方法二:本地安装(用于开发)

# 1. 通过 npm 安装(审查包后)
npm view openclaw-swarm-orchestrator  # 安装前查看
npm install -g openclaw-swarm-orchestrator

# 2. 验证安装
swarm-orchestrator --version

# 3. 启动服务(在不同终端中)

# 终端 1: 启动 Redis
redis-server

# 终端 2: 启动后端
cd backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload

# 终端 3: 启动前端
cd frontend
npm install
npm run dev

方法三:从源码安装(最透明)

# 1. 克隆并检查
git clone https://github.com/ZhenRobotics/openclaw-swarm-orchestrator.git
cd openclaw-swarm-orchestrator

# 2. 验证提交哈希(安全检查)
git log -1 --format="%H"
# 应该是: acae6e5... (或最新发布标签)

# 3. 运行前查看代码
cat backend/requirements.txt  # 检查依赖
cat package.json              # 检查 npm 依赖
cat docker-compose.yml        # 检查容器配置

# 4. 安装依赖
cd backend && pip install -r requirements.txt
cd ../frontend && npm install

# 5. 运行(详细步骤见方法二)

⚙️ 配置

最小配置(仅本地)

在项目根目录创建 .env

# 最小配置 - 无需外部服务
DATABASE_URL=sqlite+aiosqlite:///./data/swarm.db
REDIS_URL=redis://localhost:6379
SECRET_KEY=your-random-secret-key-here
DEBUG=true

可选:大模型智能体支持

如果您想使用大模型智能体(OpenAI、Anthropic),添加:

# 可选 - 仅在使用大模型智能体时需要
OPENAI_API_KEY=sk-your-key-here
ANTHROPIC_API_KEY=sk-ant-your-key-here

⚠️ 安全提示:

  • 仅在计划使用大模型智能体时添加 API 密钥
  • 安全存储 .env 文件(不要提交到 git)
  • API 密钥永远不会发送到我们的服务器(仅发送到官方大模型提供商)

💻 基本使用

1. 启动系统

# 使用 Docker
docker-compose up -d

# 或手动启动
redis-server &
cd backend && uvicorn app.main:app --reload &
cd frontend && npm run dev &

2. 创建您的第一个智能体

通过 Web UI:

  1. 打开 http://localhost:3000
  2. 转到"智能体"页面
  3. 点击"新建智能体"
  4. 填写:
    • 名称:"我的助手"
    • 类型:"llm"(或"tool"、"human"、"custom")
    • 配置:{"model": "gpt-4"}(如果使用大模型)

通过 API:

curl -X POST http://localhost:8000/api/agents \
  -H "Content-Type: application/json" \
  -d '{
    "name": "我的助手",
    "type": "llm",
    "config": {"model": "gpt-4"}
  }'

3. 创建任务

curl -X POST http://localhost:8000/api/tasks \
  -H "Content-Type: application/json" \
  -d '{
    "title": "分析数据",
    "description": "处理销售报告",
    "priority": "high"
  }'

4. 监控状态

Web 仪表板: http://localhost:3000

API:

# 系统状态
curl http://localhost:8000/api/orchestrator/status

# 列出智能体
curl http://localhost:8000/api/agents

# 列出任务
curl http://localhost:8000/api/tasks

🔧 智能体类型

1. LLM 智能体(需要 API 密钥)

使用外部大模型 API(OpenAI、Anthropic)。

{
  "name": "GPT-4 助手",
  "type": "llm",
  "config": {
    "model": "gpt-4",
    "temperature": 0.7
  }
}

需要: OPENAI_API_KEYANTHROPIC_API_KEY

2. 工具智能体(仅本地)

执行本地函数/脚本。

{
  "name": "数据处理器",
  "type": "tool",
  "config": {
    "script_path": "./tools/process_data.py"
  }
}

无需外部服务。

3. 人类智能体(仅本地)

人在回路工作流。

{
  "name": "管理员审批",
  "type": "human",
  "config": {
    "notification": "email"
  }
}

无需外部服务。

4. 自定义智能体(用户定义)

您定义行为。

from swarm_orchestrator.base import BaseAgent

class MyCustomAgent(BaseAgent):
    async def execute(self, task):
        # 您的自定义逻辑
        return result

无需外部服务。


📊 监控与日志

Web 仪表板

访问 http://localhost:3000:

  • 实时智能体状态
  • 任务队列监控
  • 系统统计
  • 执行日志

日志文件

所有日志本地存储:

# 应用日志
tail -f logs/swarm.log

# Docker 日志(如果使用 Docker)
docker-compose logs -f

🔐 安全最佳实践

1. API 密钥

  • ✅ 存储在 .env 文件中(不在代码中)
  • ✅ 设置文件权限:chmod 600 .env
  • ✅ 将 .env 添加到 .gitignore
  • ✅ 永远不要将 API 密钥提交到 git

2. 网络安全

  • ✅ 防火墙:阻止端口 8000、3000 的外部访问
  • ✅ 开发时仅使用 localhost(不是 0.0.0.0)
  • ✅ 生产环境启用 HTTPS

3. 数据隐私

  • ✅ 所有数据本地存储在 ./data/
  • ✅ 数据库文件权限:chmod 600 data/swarm.db
  • ✅ 定期备份:cp data/swarm.db backups/

🐛 故障排除

后端无法启动

# 检查 Python 版本
python --version  # 必须是 3.11+

# 检查 Redis
redis-cli ping  # 必须返回 PONG

# 检查日志
tail -f logs/swarm.log

前端无法启动

# 检查 Node 版本
node --version  # 必须是 18+

# 清除缓存
cd frontend
rm -rf node_modules
npm install

端口冲突

# 检查端口
lsof -i :8000  # 后端
lsof -i :3000  # 前端
lsof -i :6379  # Redis

# 如需终止进程
kill -9 <PID>

📚 文档


🤝 支持与社区


✅ 安装前检查清单

使用本技能前:

  • 在 GitHub 上查看源代码
  • 验证提交哈希:acae6e5
  • 检查 requirements.txtpackage.json
  • 阅读上述安全策略
  • 了解数据存储位置
  • 知道需要哪些 API 密钥(如果有)
  • 首先在隔离环境中运行(可选)

版本: 0.1.0 状态: Alpha - 活跃开发中 本地优先: ✅ 所有核心功能离线工作 隐私: ✅ 数据不离开您的机器(可选的大模型调用除外)


🇬🇧 English Version

Swarm Orchestrator Skill

Status: 🟢 Local-First AI Agent Orchestration Platform Type: Self-hosted, no external dependencies required Privacy: 100% local processing (except optional LLM API calls)


🔒 Security & Trust

What This Skill Does

  • Runs locally on your machine (backend + frontend)
  • No telemetry or data collection
  • No external servers required for core functionality
  • Optional API keys only needed if you use LLM agents (OpenAI/Claude)
  • Open source - all code is auditable

What This Skill Does NOT Do

  • ❌ Does not send data to external servers (except optional LLM APIs)
  • ❌ Does not collect analytics or telemetry
  • ❌ Does not require account registration
  • ❌ Does not access your files without permission
  • ❌ Does not run background processes without your knowledge

Data Storage

  • Database: SQLite file in ./data/swarm.db (local only)
  • Logs: Text files in ./logs/ (local only)
  • Cache: Redis on localhost (local only)
  • No cloud sync or remote storage

📋 Overview

OpenClaw Swarm Orchestrator is a local-first platform for building and managing multi-agent AI systems. Think of it as a "control tower" for coordinating multiple AI agents working together.

Core Features

  • Agent Registry - Register LLM, Tool, Human, and Custom agents
  • Task Queue - Priority-based task distribution with dependencies
  • Real-time Dashboard - Web UI to monitor agents and tasks
  • RESTful API - Complete REST API for programmatic control
  • Local Storage - All data stays on your machine

Architecture

┌─────────────────┐
│  Web Dashboard  │  http://localhost:3000
│   (React UI)    │
└────────┬────────┘
         │
┌────────▼────────┐
│  FastAPI Server │  http://localhost:8000
│  (Backend API)  │
└────────┬────────┘
         │
┌────────▼────────┐
│ Local Storage   │
│ • SQLite DB     │  ./data/swarm.db
│ • Redis Cache   │  localhost:6379
│ • Log Files     │  ./logs/*.log
└─────────────────┘

🚀 Installation

Prerequisites Check

Before installing, verify you have:

# Python 3.11+
python --version

# Node.js 18+
node --version

# Redis
redis-cli ping  # Should return PONG

# (Optional) Docker
docker --version

Method 1: Using Docker (Recommended - Easiest)

This is the safest method - everything runs in containers.

# 1. Clone repository (inspect code first!)
git clone https://github.com/ZhenRobotics/openclaw-swarm-orchestrator.git
cd openclaw-swarm-orchestrator

# 2. Review docker-compose.yml before starting
cat docker-compose.yml

# 3. Start services
docker-compose up -d

# 4. Verify
curl http://localhost:8000/health
# Should return: {"status": "healthy"}

Access:

Method 2: Local Installation (For Development)

# 1. Install via npm (after reviewing package)
npm view openclaw-swarm-orchestrator  # Review before installing
npm install -g openclaw-swarm-orchestrator

# 2. Verify installation
swarm-orchestrator --version

# 3. Start services (in separate terminals)

# Terminal 1: Start Redis
redis-server

# Terminal 2: Start backend
cd backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload

# Terminal 3: Start frontend
cd frontend
npm install
npm run dev

Method 3: From Source (Most Transparent)

# 1. Clone and inspect
git clone https://github.com/ZhenRobotics/openclaw-swarm-orchestrator.git
cd openclaw-swarm-orchestrator

# 2. Verify commit hash (security check)
git log -1 --format="%H"
# Should be: acae6e5... (or latest release tag)

# 3. Review code before running
cat backend/requirements.txt  # Check dependencies
cat package.json              # Check npm deps
cat docker-compose.yml        # Check container config

# 4. Install dependencies
cd backend && pip install -r requirements.txt
cd ../frontend && npm install

# 5. Run (see Method 2 for detailed steps)

⚙️ Configuration

Minimal Configuration (Local Only)

Create .env in project root:

# Minimal config - no external services needed
DATABASE_URL=sqlite+aiosqlite:///./data/swarm.db
REDIS_URL=redis://localhost:6379
SECRET_KEY=your-random-secret-key-here
DEBUG=true

Optional: LLM Agent Support

If you want to use LLM agents (OpenAI, Anthropic), add:

# Optional - only if using LLM agents
OPENAI_API_KEY=sk-your-key-here
ANTHROPIC_API_KEY=sk-ant-your-key-here

⚠️ Security Note:

  • Only add API keys if you plan to use LLM agents
  • Store .env file securely (not in git)
  • API keys are never sent to our servers (only to official LLM providers)

💻 Basic Usage

1. Start the System

# Using Docker
docker-compose up -d

# Or manually
redis-server &
cd backend && uvicorn app.main:app --reload &
cd frontend && npm run dev &

2. Create Your First Agent

Via Web UI:

  1. Open http://localhost:3000
  2. Go to "Agents" page
  3. Click "New Agent"
  4. Fill in:
    • Name: "My Assistant"
    • Type: "llm" (or "tool", "human", "custom")
    • Config: {"model": "gpt-4"} (if using LLM)

Via API:

curl -X POST http://localhost:8000/api/agents \
  -H "Content-Type: application/json" \
  -d '{
    "name": "My Assistant",
    "type": "llm",
    "config": {"model": "gpt-4"}
  }'

3. Create a Task

curl -X POST http://localhost:8000/api/tasks \
  -H "Content-Type: application/json" \
  -d '{
    "title": "Analyze data",
    "description": "Process the sales report",
    "priority": "high"
  }'

4. Monitor Status

Web Dashboard: http://localhost:3000

API:

# System status
curl http://localhost:8000/api/orchestrator/status

# List agents
curl http://localhost:8000/api/agents

# List tasks
curl http://localhost:8000/api/tasks

🔧 Agent Types

1. LLM Agents (Requires API Key)

Uses external LLM APIs (OpenAI, Anthropic).

{
  "name": "GPT-4 Assistant",
  "type": "llm",
  "config": {
    "model": "gpt-4",
    "temperature": 0.7
  }
}

Required: OPENAI_API_KEY or ANTHROPIC_API_KEY

2. Tool Agents (Local Only)

Executes local functions/scripts.

{
  "name": "Data Processor",
  "type": "tool",
  "config": {
    "script_path": "./tools/process_data.py"
  }
}

No external services needed.

3. Human Agents (Local Only)

Human-in-the-loop workflows.

{
  "name": "Manager Approval",
  "type": "human",
  "config": {
    "notification": "email"
  }
}

No external services needed.

4. Custom Agents (User-Defined)

You define the behavior.

from swarm_orchestrator.base import BaseAgent

class MyCustomAgent(BaseAgent):
    async def execute(self, task):
        # Your custom logic
        return result

No external services needed.


📊 Monitoring & Logs

Web Dashboard

Access at http://localhost:3000:

  • Real-time agent status
  • Task queue monitoring
  • System statistics
  • Execution logs

Log Files

All logs stored locally:

# Application logs
tail -f logs/swarm.log

# Docker logs (if using Docker)
docker-compose logs -f

🔐 Security Best Practices

1. API Keys

  • ✅ Store in .env file (not in code)
  • ✅ Set file permissions: chmod 600 .env
  • ✅ Add .env to .gitignore
  • ✅ Never commit API keys to git

2. Network Security

  • ✅ Firewall: Block ports 8000, 3000 from external access
  • ✅ Use localhost only (not 0.0.0.0) for development
  • ✅ Enable HTTPS in production

3. Data Privacy

  • ✅ All data stored locally in ./data/
  • ✅ Database file permissions: chmod 600 data/swarm.db
  • ✅ Regular backups: cp data/swarm.db backups/

🐛 Troubleshooting

Backend won't start

# Check Python version
python --version  # Must be 3.11+

# Check Redis
redis-cli ping  # Must return PONG

# Check logs
tail -f logs/swarm.log

Frontend won't start

# Check Node version
node --version  # Must be 18+

# Clear cache
cd frontend
rm -rf node_modules
npm install

Port conflicts

# Check ports
lsof -i :8000  # Backend
lsof -i :3000  # Frontend
lsof -i :6379  # Redis

# Kill processes if needed
kill -9 <PID>

📚 Documentation


🤝 Support & Community


✅ Pre-Installation Checklist

Before using this skill:

  • Review source code on GitHub
  • Verify commit hash: acae6e5
  • Check requirements.txt and package.json
  • Read security policy above
  • Understand data storage locations
  • Know which API keys you need (if any)
  • Run in isolated environment first (optional)

Version: 0.1.0 Status: Alpha - Active Development Local-First: ✅ All core features work offline Privacy: ✅ No data leaves your machine (except optional LLM calls)


Built with privacy and transparency in mind. Inspect the code before you trust it.

Source Transparency

This detail page is rendered from real SKILL.md content. Trust labels are metadata-based hints, not a safety guarantee.

Related Skills

Related by shared tags or category signals.

Security

Network AI

Python orchestration skill: local multi-agent workflows via blackboard file, permission gating, and token budget scripts. All execution is local — no network...

Registry SourceRecently Updated
61.4K
Profile unavailable
Automation

Multi-Agent Orchestrator

Production-grade multi-agent orchestration patterns. Decompose complex tasks into parallel subtasks, coordinate agent swarms, build sequential pipelines, and...

Registry SourceRecently Updated
098
Profile unavailable
Automation

Agent Orchestrator

Multi-agent orchestration with 5 proven patterns - Work Crew, Supervisor, Pipeline, Council, and Auto-Routing

Registry SourceRecently Updated
0582
Profile unavailable