claude-flow (ruflo)

★ 64.160

Multi-agent orchestration for Claude Code, Codex & co: swarms, adaptive memory, RAG integration.

claude-flow (repo renamed to ruflo; the npm packages claude-flow and ruflo exist in parallel, currently both at version 3.25.6) is a multi-agent orchestration framework for Claude Code, Codex, and other coding agents. Its centerpiece is a "queen-led" swarm coordination: multiple agents work on subtasks, synchronize via consensus mechanisms (Raft, Byzantine, Gossip), and share a common memory.

This memory (AgentDB) is HNSW-indexed for vector search and, according to the project docs, several times faster than a naive brute-force search on smaller datasets; hybrid search and graph traversal are also included. Via SONA neural patterns and a "ReasoningBank", agents are meant to learn from past task trajectories and reuse successful strategies on similar future tasks.

The framework is model-agnostic (Claude, GPT, Gemini, Cohere, local via Ollama) with automatic routing, ships around 210 MCP tools across five server groups (Core, Intelligence, Agents, Memory, DevTools), and can also be installed as a set of Claude Code plugins (including ruflo-core, ruflo-swarm, ruflo-rag-memory). A federation feature lets you network multiple claude-flow instances across team or organization boundaries via WebSocket and hand off tasks between them.

🔧 SETUP

Interactive setup: `npx ruflo@latest init wizard` (without the wizard: `npx ruflo@latest init`)
Alternatively install globally: `npm install -g ruflo@latest`
Register as an MCP server with Claude Code: `claude mcp add ruflo -- npx ruflo@latest mcp start`
Optionally as Claude Code plugins: `/plugin install ruflo-core@ruflo`, `/plugin install ruflo-swarm@ruflo`, `/plugin install ruflo-rag-memory@ruflo`
For cross-team networking: `npx claude-flow@latest federation init` or `federation join wss://...`

✅ WHEN TO USE IT?

Coordinating multiple agents in parallel

A complex feature is broken down into subtasks, and multiple agents work on it simultaneously via a swarm topology, sharing a common memory.

Agents should learn from experience

SONA/ReasoningBank remembers successful task trajectories and reapplies them to similar future tasks.

⛔ WHEN NOT TO?

A single, simple task

The overhead of swarm, memory, and consensus mechanisms doesn't pay off for a single, clearly scoped task that one agent can solve alone.

Better: Use Claude Code directly, without an orchestration layer.

Cross-team federation without infrastructure

Federation needs a reachable WebSocket endpoint (wss://) — without your own server infrastructure it's not practically usable.

Better: Run a local setup without federation commands, just a single instance.