☁️🐾 CloudClaw

Enterprise-grade Open Source AI Agent Platform

Build, deploy, and scale AI Agents with Spring Boot & Spring AI

English · 中文 · Quick Start · Documentation

---

🌟 Why CloudClaw

Most AI Agent platforms target individual developers, relying on local filesystems and script execution. Enterprises need something different: multi-tenant isolation, stateless scalability, and a secure Agent runtime.

CloudClaw exists to:

• 🚀 Make AI Agent platforms as easy to deploy as any regular app — java -jar to start, Docker Compose for production
• 👥 Multi-tenant by design — Sessions, memories, and configs are isolated per user, not a single-user toy
• ⚡ Stateless Agents — All state lives in databases and caches, enabling horizontal scaling
• 🔒 Security first — Agent I/O through MCP, database, or isolated sandbox — never raw host access
• 🧩 Pluggable architecture — Memory engines, message queues, LLM providers, and MCP servers are all replaceable

✨ Features

• 🤖 Multi-Agent — Create and manage multiple AI agents with independent configs
• 🔧 MCP Integration — Built-in MCP gateway with connection pooling, tool routing, and permission control
• 🧠 Memory Service — Pluggable memory engine (JDBC / Mem0) with vector search support
• 💬 Session Management — Session persistence, context caching, message history
• 🔐 Auth & RBAC — JWT authentication with admin/user role separation
• 🎯 Skill System — Claude Agent Skill standard with progressive disclosure (metadata → instructions → files), auto-discovery via @Tool
• 📡 Message Queue — Redis Streams or in-memory MQ for async processing
• 🖥️ Dual UI — Admin dashboard + Chat interface in one Vue.js app
• 🐘 PostgreSQL — Cluster mode with Redis cache
• 🪶 Standalone Mode — SQLite, zero external dependencies
• 📊 Monitoring — System logs, prompt logs, usage dashboards
• 🔒 Sandbox Execution — Isolated code execution (Python/JS/Shell/Java) with Local, Docker, and E2B backends, Stateless and Session modes
• 🧩 LLM Management — Multi-provider credential management, model registry, usage tracking

🔄 Multi-Agent Workflow

CloudClaw supports 5 built-in orchestration modes for multi-agent collaboration:

Mode	Description
Pipeline	Sequential steps — output of one node feeds into the next
Parallel	Multiple agents run simultaneously, results merged (concat or LLM summarize)
Router	LLM-based intent classification routes to the best-matching sub-agent
Supervisor	A planner/reviewer agent delegates tasks to specialist sub-agents iteratively
Handoff	Conversation is transferred between agents, each maintaining its own context

All modes are configured inline (no external agent references needed) and support per-node model, system prompt, MCP servers, and skills.

🆚 CloudClaw vs OpenClaw

	OpenClaw	CloudClaw
Target	Personal AI assistant	Enterprise AI Agent platform
Language	Node.js / TypeScript	Java (Spring Boot)
Users	Single user	Multi-tenant with data isolation
Storage	Local filesystem (MEMORY.md)	Database (PostgreSQL / SQLite)
Script Execution	Local Shell / PTY	Isolated Sandbox (Docker / E2B)
Agent State	Stateful (local process)	Stateless, horizontally scalable
Memory	Local Markdown files	Database-backed (JDBC / Mem0)
AI Framework	Custom	Spring AI
Frontend	None (third-party integrations)	Vue 3 + Element Plus (Admin + Chat)
Deployment	Personal devices	Server / Container / K8s
License	MIT	Apache 2.0

OpenClaw is your personal butler. CloudClaw is your enterprise Agent middleware.

🏗 Architecture

cloudclaw
├── cloudclaw-app          # Spring Boot entry point & configuration
├── cloudclaw-common       # Shared models, DTOs, utilities
├── cloudclaw-auth         # JWT authentication & authorization
├── cloudclaw-agent        # Agent engine, prompt assembly, chat orchestration
├── cloudclaw-llm          # LLM multi-provider routing, credential encryption, usage stats
├── cloudclaw-mcp          # MCP gateway, connection pool, tool routing
├── cloudclaw-memory       # Memory service (JDBC / Mem0 engines)
├── cloudclaw-session      # Session & message persistence
├── cloudclaw-skill        # Skill definition & management
├── cloudclaw-mq           # Message queue abstraction (Redis Streams / in-memory)
├── cloudclaw-admin        # Admin API controllers
├── cloudclaw-user         # User-facing API controllers
├── cloudclaw-sandbox      # Code sandbox (Local/Docker/E2B, Stateless/Session modes)
├── cloudclaw-standalone   # Standalone mode (SQLite, in-memory MQ, Caffeine cache)
└── cloudclaw-ui           # Vue.js frontend (Chat + Admin merged)

🚀 Quick Start

Option 1: Download Release (Zero Dependencies)

# Download from GitHub Releases
wget https://github.com/cloudclaw-dev/cloudclaw/releases/download/v1.0.1/cloudclaw-1.0.1-release.zip
unzip cloudclaw-1.0.1-release.zip

# Start
chmod +x start.sh
./start.sh

# Access http://localhost:8080
# Default login: admin / admin123

Option 2: Build from Source

git clone https://github.com/cloudclaw-dev/cloudclaw.git
cd cloudclaw
mvn clean package -DskipTests
java -jar cloudclaw-app/target/cloudclaw-app-1.0.1.jar

Option 3: Docker Compose

# Standalone
docker compose -f docker-compose.standalone.yml up -d

# Cluster (PostgreSQL + Redis)
docker compose -f docker-compose.yml up -d

Cluster Mode Configuration

# Prerequisites: PostgreSQL 16 + Redis 7
createdb cloudclaw

java -jar cloudclaw-app/target/cloudclaw-app-1.0.1.jar \
  --spring.profiles.active=cluster

Configure application-cluster.yml:

spring:
  datasource:
    url: jdbc:postgresql://localhost:5432/cloudclaw
    username: postgres
    password: your-password
  data:
    redis:
      host: localhost
      port: 6379

After Login

Go to LLM 管理 to configure your API key and model provider. Supports OpenAI, DeepSeek, Qwen, GLM, Ollama, and any OpenAI-compatible API.

⚙️ Configuration

Property	Default	Description
spring.profiles.active	standalone	standalone or cluster
cloudclaw.jwt.secret	(built-in)	JWT signing key (change in production!)
cloudclaw.jwt.access-token-ttl	2h	Access token TTL
cloudclaw.jwt.refresh-token-ttl	7d	Refresh token TTL
cloudclaw.memory.engine	jdbc	Memory engine: jdbc or mem0
cloudclaw.mq.provider	inmemory	MQ provider: inmemory or redis
cloudclaw.mcp.pool.max-connections-per-server	5	Max connections per MCP server
cloudclaw.cache.session-ttl	30m	Session cache TTL
cloudclaw.sandbox.default-backend	LOCAL	Sandbox backend: LOCAL, DOCKER, E2B
cloudclaw.sandbox.default-mode	STATELESS	Sandbox mode: STATELESS or SESSION
cloudclaw.sandbox.default-timeout	30s	Default code execution timeout
cloudclaw.sandbox.max-timeout	5m	Maximum allowed execution timeout

See CONTRIBUTING.md for development setup.

📦 Deployment Modes

Mode	Database	MQ	Cache	Best For
Standalone	SQLite	In-Memory	Caffeine	Dev, testing, personal use
Cluster	PostgreSQL	Redis Streams	Redis	Production, horizontal scaling

🛠 Tech Stack

Layer	Technology
Runtime	Java 17, Spring Boot 3.4.5
AI Framework	Spring AI 1.1.5
Database	PostgreSQL 16 / SQLite
Cache & MQ	Redis 7 / Caffeine / In-Memory
Auth	JWT + Spring Security
Migration	Flyway
Sandbox	agent-sandbox-core, Docker (Testcontainers), E2B
Frontend	Vue 3, Element Plus, Vite, TypeScript, ECharts

📄 License

Apache License 2.0

---

⬆ Back to Top

Made with ❤️ by the CloudClaw Team