Ruflo (formerly Claude Flow) is an agent meta-harness that wraps Claude Code and Codex with multi-agent coordination, persistent memory, and routing. It adds more than 100 specialized agents for coding, testing, security, documentation, and architecture, and organizes them into swarms that share memory and reach decisions by consensus. The conceptual flow runs from the user through the Ruflo CLI or MCP server to a router, a swarm, its agents, memory, and finally the LLM providers, with memory feeding past experience back into routing.
There are two ways to use it. Installed as Claude Code plugins, it adds slash commands, skills, and agent definitions without writing files into the workspace. The full CLI setup adds the complete loop: agents, commands, skills, an MCP server, hooks, and a background daemon, after which hooks route tasks and coordinate agents while you keep using Claude Code as normal. Any MCP client, Codex included, can connect to the Ruflo MCP server.
A plugin marketplace splits the functionality into packages for orchestration, memory and knowledge graphs, self-learning, test generation, browser testing, security scanning, architecture methods (ADR, DDD, and SPARC), observability, and cost tracking, plus domain plugins for IoT and trading. A console plugin gives Claude Code a cockpit for missions and swarm status, with a control that pauses Claude's console tools.
| Capability | Claude Code alone | With Ruflo |
|---|---|---|
| Agent collaboration | Isolated, no shared context | Swarms with shared memory and consensus |
| Coordination | Manual orchestration | Queen-led hierarchy (Raft, Byzantine, Gossip) |
| Memory | Session-only | HNSW vector memory |
| Background workers | None | 12 auto-triggered workers |
| LLM providers | Anthropic only | 5 providers with failover |
Features
- Swarm coordination: hierarchical, mesh, and adaptive topologies with consensus across agents
- Vector memory: HNSW-indexed AgentDB that keeps context across sessions, with hybrid search and graph hops for retrieval
- Self-learning: SONA neural patterns, ReasoningBank, and trajectory learning that feed successful patterns back into routing
- Agent federation: agents on different machines or organizations discover and authenticate each other through mTLS and ed25519, with PII filtering and audit trails
- Background workers: 12 auto-triggered workers for tasks such as audits, optimization, and test-gap detection
- Plugin marketplace: 33 Claude Code plugins and 21 npm plugins, browsable and installable from the console
- Multi-provider routing: Claude, GPT, Gemini, Cohere, and Ollama models with routing and failover
- MetaHarness: grades an agent setup's readiness, scans tool configs for security issues, and snapshots the project to catch regressions
- Security: AIDefence prompt-injection blocking, PII detection, input validation, and CVE scanning
- Console: a Claude Code cockpit for creating missions, viewing swarm status, and taking back control from Claude
