Claude Flow

Claude Flow (also known as Ruflo v3) is an enterprise-grade multi-agent AI orchestration framework that transforms Claude Code into a powerful development platform. It enables teams to deploy, coordinate, and optimize 60+ specialized AI agents working together on complex software engineering tasks through swarm coordination, self-learning intelligence, and fault-tolerant consensus protocols.

• Swarm Coordination organizes agents into configurable topologies including hierarchical (queen-led), mesh (peer-to-peer), ring (sequential pipeline), and star (hub-spoke) patterns. Queens coordinate workers, prevent goal drift, and reach consensus on decisions using five distributed protocols: Raft, Byzantine Fault Tolerance (handles up to one-third faulty agents), Gossip, CRDT, and Quorum.

• 60+ Specialized Agents cover coding, code review, testing, security audits, documentation, DevOps, architecture, and performance analysis. Agents are organized across eight categories including core development, swarm coordination, consensus, GitHub integration, and SPARC methodology.

• Self-Learning Intelligence (SONA) uses a Self-Optimizing Neural Architecture that learns which agents perform best for each task type, routing work with under 0.05ms adaptation time. The system includes nine reinforcement learning algorithms (PPO, DQN, SARSA, and others), Elastic Weight Consolidation to prevent forgetting successful patterns, and a Mixture of Experts router with eight specialized expert networks.

• HNSW Vector Memory provides 150x to 12,500x faster pattern retrieval through Hierarchical Navigable Small World indexing, with persistent storage via AgentDB and SQLite. A knowledge graph using PageRank and community detection identifies influential insights across sessions.

• Intelligent Task Routing analyzes task complexity and automatically routes work across three tiers: Agent Booster (WebAssembly) handles simple code transforms at under 1ms with zero LLM cost, medium tasks go to faster models like Haiku or Sonnet, and complex architecture decisions use Opus. This extends Claude Code usage by up to 2.5x.

• Multi-Provider LLM Support works with Anthropic (Claude), OpenAI (GPT), Google (Gemini), xAI (Grok), Mistral, and local models via Ollama, with automatic failover and cost-based routing that can reduce API costs by up to 85 percent.

• MCP Integration provides native Model Context Protocol support with 170+ tools across coordination, monitoring, memory, GitHub, and security categories. Compatible with Claude Code, Claude Desktop, VS Code, Cursor, Windsurf, JetBrains, and ChatGPT.

• Hive Mind System implements queen-led collective intelligence with three queen types (Strategic, Tactical, Adaptive) directing eight worker specializations through collective memory with LRU caching and SQLite persistence.

• Security (AIDefence) detects prompt injection, jailbreak attempts, and PII exposure in under 10ms with 50+ detection patterns, input validation, path traversal prevention, and command sandboxing.

• Context Autopilot eliminates context window limitations by automatically archiving conversation turns to SQLite, blocking premature compaction, and restoring importance-ranked context after resets.

• Extensible Plugin System supports custom capabilities through a plugin SDK with WebAssembly acceleration, shared via a decentralized IPFS marketplace with Ed25519 signature verification.

To get started, install via npx ruflo@alpha init or use the one-line curl installer. Add as an MCP server to Claude Code with claude mcp add ruflo -- npx ruflo@latest mcp start. The framework requires Node.js 20+ and supports Windows, macOS, and Linux.