Key takeaways
- Orchestrates 60+ specialized agents (coder, tester, reviewer, architect, security, etc.) in coordinated swarms on top of Claude Code
- Supports multiple swarm topologies: mesh, hierarchical, ring, and star — with Raft, BFT, Gossip, and CRDT consensus protocols
- Self-learning architecture with Q-Learning router, Mixture of Experts (8), 42+ skills, and a RuVector intelligence layer for continuous optimization
- Native MCP protocol support — can run as an MCP server, integrating with Claude Code and other MCP-compatible tools
- Memory layer (AgentDB) with vector search (HNSW), flash attention, and elastic weight consolidation to prevent catastrophic forgetting
- Multi-provider: works with Claude, GPT, Gemini, and Ollama
FAQ
What is Claude Flow?
Claude Flow (rebranded as Ruflo v3) is an open-source multi-agent orchestration platform that turns Claude Code into a swarm-based development system with 60+ specialized agents.
What swarm topologies does it support?
Mesh, hierarchical, ring, and star topologies, with consensus via Raft, BFT, Gossip, and CRDT protocols.
How does memory work?
AgentDB provides persistent memory with HNSW vector search (150x–12,500x faster retrieval), flash attention, and elastic weight consolidation to prevent forgetting across sessions.
Does it support MCP?
Yes. It runs as an MCP server and integrates natively with Claude Code and other MCP-compatible tools.
Who built it?
Created by Reuven Cohen (ruvnet). MIT licensed with 14k+ GitHub stars and 1.7k forks.
Overview
Claude Flow (now Ruflo v3) is an open-source agent orchestration platform that transforms Claude Code into a multi-agent development system. It coordinates 60+ specialized agents — coders, testers, reviewers, architects, security auditors — working together on complex software engineering tasks.
Architecture
The system follows a layered architecture:
- Entry Layer — CLI and MCP server with AI-based security (AIDefence)
- Routing Layer — Q-Learning router with 8 Mixture of Experts, 42+ skills, and 17 hooks
- Swarm Coordination — Configurable topologies (mesh, hierarchical, ring, star) with consensus protocols (Raft, BFT, Gossip, CRDT) and human-agent coordination via Claims
- Agent Layer — 60+ specialized agents dispatched based on task type
- Resource Layer — Memory (AgentDB), multi-provider LLM support, and 12 background workers
RuVector Intelligence Layer
The self-learning subsystem includes:
- SONA — Self-optimization with sub-0.05ms latency
- EWC++ — Elastic weight consolidation to prevent catastrophic forgetting
- Flash Attention — 2.49–7.47x speedup on attention operations
- HNSW — Vector search with 150x–12,500x faster retrieval
- ReasoningBank — Pattern storage for reuse across tasks
- 9 RL algorithms — Q-Learning, SARSA, PPO, DQN, and more for continuous routing improvement
Key Differentiators
- Swarm-native: Not just multi-agent — true swarm coordination with consensus and topology selection
- Self-learning loop: RETRIEVE → JUDGE → DISTILL → CONSOLIDATE → ROUTE cycle continuously improves agent routing
- MCP-first: Built as an MCP server, making it composable with the broader Claude ecosystem
- Multi-provider: Works with Claude, GPT, Gemini, and Ollama — not locked to Anthropic
Competitive Context
Claude Flow occupies the "swarm orchestration" niche — heavier than simple multi-agent frameworks like CrewAI, more research-oriented than production platforms like Tembo. The 14k stars signal strong community interest in autonomous agent swarms, though the complexity of the system (consensus protocols, RL routing, vector intelligence layers) raises questions about real-world production adoption vs. experimental usage.