🪴 Brain

Ruflo

4 min read

Ruflo

Multi-agent orchestration platform for Claude Code and other harnesses. Claude Flow rebranded as Ruflo by rUv — Rust engine, embeddings, memory, plugin system from Cognitum.One. 100+ specialized agents coordinate in swarms, learn from every task, remember across sessions, and federate across trust boundaries.

🚀 What it is

  • Repo: https://github.com/ruvnet/ruflo
  • CLI: npx ruflo@latest init wizard (cross-platform) or npm install -g ruflo@latest
  • MCP server: claude mcp add ruflo -- npx ruflo@latest mcp start
  • Full install: 98 agents, 60+ commands, 30 skills, MCP server, hooks, daemon

🧩 Two installation paths (#1744)

Claude Code PluginCLI install (npx ruflo init)
What you getSlash commands + a few skills + agent definitionsFull loop — 98 agents, 60+ commands, 30 skills, MCP, hooks, daemon
Files in workspaceNone.claude/, .claude-flow/, CLAUDE.md, helpers
MCP serverNo (memory_store, swarm_init unavailable)Yes
HooksNoYes
Best forTrying a single pluginProduction — everything works as in the docs

🐝 What you get

CapabilityDescription
100+ agentsCoding, testing, security, docs, architecture
Comms LayerZero-trust federation — mTLS + ed25519, PII-gated, behavioral trust scoring
Swarm CoordinationHierarchical, mesh, adaptive topologies — Raft, Byzantine, Gossip consensus
Self-LearningSONA neural patterns, ReasoningBank, trajectory learning
Vector MemoryHNSW-indexed AgentDB — 150x–12,500x faster than brute force, sub-ms retrieval
Background Workers12 auto-triggered (audit, optimize, testgaps, …)
Plugins32 native Claude Code plugins + 21 npm
Multi-Provider LLMClaude, GPT, Gemini, Cohere, Ollama — smart routing with failover
SecurityAIDefence (prompt injection, PII detection), CVE remediation, path traversal prevention
Task RoutingIntelligent (89% accuracy)

🔌 32 plugins (selection)

Core/Orchestration: ruflo-core · ruflo-swarm · ruflo-autopilot · ruflo-loop-workers · ruflo-workflows · ruflo-federation

Memory/Knowledge: ruflo-agentdb · ruflo-rag-memory · ruflo-rvf · ruflo-ruvector (GPU + Graph RAG, 103 tools) · ruflo-knowledge-graph

Intelligence/Learning: ruflo-intelligence · ruflo-daa · ruflo-ruvllm (local LLMs with routing) · ruflo-goals

Code Quality/Testing: ruflo-testgen · ruflo-browser (Playwright) · ruflo-jujutsu (git diff risk scoring) · ruflo-docs

Security: ruflo-security-audit · ruflo-aidefence

Methodology: ruflo-adr · ruflo-ddd · ruflo-sparc (5-phase methodology with quality gates)

DevOps: ruflo-migrations · ruflo-observability · ruflo-cost-tracker

Extensibility: ruflo-agent (WASM sandbox + Anthropic Managed Agents) · ruflo-plugin-creator

Domain-specific: ruflo-iot-cognitum · ruflo-neural-trader (112+ tools) · ruflo-market-data

☘️ What changes vs plain Claude Code

CapabilityClaude Code alone+ Ruflo
Agent collabIsolated, no shared contextSwarms with shared memory + consensus
CoordinationManual orchestrationQueen-led hierarchy (Raft/Byzantine/Gossip)
MemorySession-onlyHNSW vector, sub-ms retrieval
LearningStaticSONA self-learning, pattern matching
Task routingYou decideIntelligent (89% accuracy)
Background workersNone12 auto-triggered
LLM providersAnthropic only5 providers with failover
SecurityStandardCVE-hardened + AIDefence

💬 Web UI — flo.ruv.io

Multi-model AI chat with native MCP tool calling:

  • 6 frontier models out of the box (Qwen 3.6 Max default, Sonnet 4.6, Haiku 4.5, Gemini 2.5 Pro/Flash, OpenAI) via OpenRouter
  • ruvLLM self-improving local model layer (MicroLoRA adapters + SONA)
  • ~210 tools (5 MCP groups) + 18-tool in-browser WASM gallery
  • Parallel tool execution (4–6+ per response)
  • Memory (“remember my favorite color is indigo” → recall weeks later)
  • Self-hostable (Docker, ruflo/src/ruvocal/Dockerfile, embedded Mongo, Cloud Run/Fly/K8s)
  • Zero install — https://flo.ruv.io/

🎯 Goal Planner — goal.ruv.io

GOAP (Goal-Oriented Action Planning) with A* search:

  • Plain-English goals → action tree with preconditions/effects/replanning
  • Live agent dashboard goal.ruv.io/agents (role, current step, memory namespace, token budget)
  • Each action node maps to an MCP tool call (~210 tools + custom)
  • Adaptive replanning when an action fails (failures = learning, not a loop)
  • Self-hostable: v3/goal_ui/ (Vite + Supabase)

🤝 Federation — “Slack for agents”

Cross-installation collaboration with zero-trust security:

  • Identity: mTLS + ed25519 challenge-response (no API keys, no shared secrets)
  • PII-gated data flow — 14-type detection, per-trust-level: BLOCK/REDACT/HASH/PASS
  • Behavioral trust scoring: 0.4×success + 0.2×uptime + 0.2×threat + 0.2×integrity — upgrades require history, downgrades are immediate
  • Compliance built in: HIPAA/SOC2/GDPR audit trails
  • 9 MCP tools + 10 CLI commands (federation_init, federation_send, federation_trust, federation_audit)
  • Opt-in WireGuard mesh (ADR-111) — packet-layer reachability tied to federation trust
npx claude-flow@latest federation init
npx claude-flow@latest federation join wss://team-b.example.com:8443
npx claude-flow@latest federation send --to team-b --type task-request --message "..."
npx claude-flow@latest federation status

✍️ Quick Start

# POSIX (macOS/Linux/WSL/Git-Bash)
curl -fsSL https://cdn.jsdelivr.net/gh/ruvnet/ruflo@main/scripts/install.sh | bash
 
# All platforms (PowerShell, cmd, …)
npx ruflo@latest init wizard
# or: npx ruflo@latest init (non-interactive)
# or: npm install -g ruflo@latest
 
# MCP in Claude Code
claude mcp add ruflo -- npx ruflo@latest mcp start

After init you use Claude Code normally — the hooks system automatically routes tasks, learns from successful patterns, and coordinates agents in the background. You don’t have to learn 314 MCP tools or 26 CLI commands.

🧩 Related

🔗 Resources

Graph View

Backlinks