hub
ruvnet Ecosystem: A Field Guide
The ruvnet ecosystem mapped: tools, repos, methods. The ideas shaping multi-agent Claude Code workflows.
Field Guide
The ruvnet Stack: Supercharging Claude Code
10 open-source projects from Reuven Cohen that transform Claude Code from a coding assistant into an enterprise agent platform. This is the missing infrastructure layer: persistent memory, coordinated swarms, deterministic retrieval, and structured builds.
The Builder
Who is ruvnet (Reuven Cohen)
Agentic Engineer and aiCTO based in Toronto. 59K+ LinkedIn followers, founder of the Agentics Foundation. Built Claude Flow on a 20-hour flight from Toronto to Bangalore. His philosophy: design systems that think, adapt, and act with bounded autonomy.
LLMs are a dead end. Not because they're useless, but because they put intelligence in the wrong place.
Everyone is calling everything an agent. An agent is defined by whether it can perceive, reason, act, and update.
Cohen built tooling out of necessity on production projects, not theory. Every repo in the ecosystem solves a real gap he hit: memory dying between sessions, agents stepping on each other, prompts burning budget unnecessarily, builds with no structure.
Priority Tier 1
Install Now: The Core Three
These three repos form the foundation. Start here before anything else.
Claude Flow / Ruflo v3
50K+ Stars · 500K DownloadsEnterprise multi-agent orchestration for Claude Code. 60+ specialized agents, 8 topology types, Byzantine fault-tolerant consensus. This is the operating system layer on top of Claude: it coordinates agents, persists memory across sessions, and self-optimizes over time.
7-Layer Architecture
User Input | v CLI Interface | v MCP + Security Layer | v Router + Skills Engine | v Swarm Orchestration (8 topologies) | v Memory (HNSW) + LLM Providers (6) | v SONA Self-Optimization
Install
npx ruflo@alpha init --wizard claude mcp add ruflo -- npx ruflo@v3alpha mcp start
| Capability | Vanilla Claude | Claude Flow |
|---|---|---|
| Agent model | Isolated sessions | Coordinated swarms |
| Memory | Session-only | Persistent HNSW |
| Optimization | Static | Self-optimizing (SONA) |
| LLM providers | Anthropic only | 6 providers |
SPARC Methodology
Anti-Vibe-CodingStructured development for agentic builds. Prevents the chaos of 'hey claude build me a CRM' and gives every build a predictable lifecycle. SPARC is not a framework. It's a discipline.
The 5 Steps
Install & Usage
pip install sparc sparc -m "Build client onboarding system" --hil sparc -m "Analyze workflow" --research-only sparc -m "Build feature" --cowboy-mode
SAFLA (Self-Aware Feedback Loop Algorithm)
Persistent MemoryPersistent memory and self-learning for agents. A 4-layer hybrid memory system that survives sessions and improves over time. 172,000 ops/sec with 60% compression.
Semantic similarity search across all stored knowledge. Finds concepts, not just keywords.
Event sequences with temporal context. Remembers when things happened and in what order.
Knowledge graphs of concepts and relationships. Maps client to engineer to stack to blockers.
Active context via attention mechanisms. Like RAM: fast access to what matters right now.
pip install safla claude mcp add safla python3 safla_mcp_enhanced.pygithub.com/ruvnet/SAFLA
Priority Tier 2
High Value: Extend the Stack
Install after the core three are running. Each solves a specific gap in cost, search, or prompt efficiency.
FACT
42ms RetrievalDeterministic retrieval at 42ms vs 2-5 seconds for traditional RAG. Facts, not hallucinations. Plugs directly into Claude as an MCP server.
github.com/ruvnet/FACTAgentic Flow
60% Cost Savings66 agents + intelligent LLM router. Routes simple tasks to cheap models, complex ones to powerful ones. 213 MCP tools included.
github.com/ruvnet/agentic-flowRuVector
Self-Learning DBSelf-learning vector database with graph fusion. Builds understanding as it indexes, not just storage. Pairs with SAFLA for complete memory.
github.com/ruvnet/ruvectorSynthLang
70% Token ReductionMath-structured prompt optimization using set theory and category theory. 30-70% cost reduction on prompts without changing intent.
github.com/ruvnet/SynthLangThe Big Picture
Full Stack Architecture
How all the pieces connect when you run the complete ruvnet stack on top of Claude Code.
Claude Code (base) | +-- + Agentic Flow MCP --> cost routing, 66 agents, 213 tools | +-- + Claude Flow MCP --> 60+ specialized agents, swarms, SPARC | +-- + SAFLA MCP --> persistent memory, self-learning | +-- + FACT --> 42ms deterministic retrieval | +-- + RuVector --> semantic search, knowledge graph | +-- + SynthLang --> 30-70% token reduction on prompts | = Sessions with memory, cheap routing, parallel swarms, fast search
Before / After
The Gap This Fills
Vanilla Claude Code is powerful. The ruvnet stack makes it production-grade.
| Gap | Without ruvnet | With ruvnet |
|---|---|---|
| Agent memory | Cold start every session | Persistent 4-layer memory |
| Multi-agent | One agent at a time | Coordinated swarms |
| Search | grep (keyword matching) | Semantic + graph (concept) |
| Cost | Always Sonnet pricing | Auto-routed by complexity |
| Build structure | Vibe coding | SPARC methodology |
| Prompt efficiency | Raw text prompts | 30-70% optimized |
Reference
All 10 Repositories
Quick-reference for the full ecosystem in install order.
| Project | Description | Install |
|---|---|---|
| Claude Flow | Multi-agent orchestration, 60+ agents, swarms | npx ruflo@alpha init --wizard |
| SPARC | Structured development methodology | pip install sparc |
| SAFLA | Persistent 4-layer memory + self-learning | pip install safla |
| FACT | 42ms deterministic retrieval | pip install fact-retrieval |
| Agentic Flow | 66 agents, intelligent LLM router, 60% savings | pip install agentic-flow |
| RuVector | Self-learning vector DB with graph fusion | pip install ruvector |
| SynthLang | Math-structured prompt optimization | pip install synthlang |
| SPARC-IDE | Custom VSCode fork for SPARC workflow | github.com/ruvnet/sparc-ide |
| Agentics | Agentics Foundation reference implementations | github.com/ruvnet/agentics |
| Bot Generator | Rapid bot prototyping and deployment | github.com/ruvnet/bot-generator |
AY Framework Integration
Where ruvnet's Ideas Fit
The AY Automate framework already uses agents, hooks, and context engineering. The ruvnet stack plugs directly into the gaps.
Agent Swarms: Claude Flow's 8 topology types map directly onto AY's Plan → Build → Review pipeline. Run architect, coder, and reviewer agents in parallel instead of sequentially.
Persistent Memory: SAFLA replaces the need to re-paste client context every session. Map client → stack → blockers once. Every subsequent session picks up from there.
SPARC on every build: Replace vibe-coding client deliverables with SPARC. Spec → Pseudocode → Architecture before a single line of code. Especially powerful on the /go skill cycle.
Cost routing: Agentic Flow's LLM router pairs with AY's multi-model strategy. Route quick summaries to Haiku, complex reasoning to Opus. Reduces client project cost significantly.
Knowledge retrieval: FACT replaces grep-style second-brain lookups with 42ms semantic search. Plug into the AY second brain for instant, accurate retrieval across all project context.
Prompt efficiency: SynthLang reduces token spend on long system prompts and CLAUDE.md content. Especially valuable for the complex agent hooks and multi-step skill definitions.
Want this running in your stack?
AY Automate builds AI automation systems for production teams.