Claude Code's swarm mode: 100 agents, one nervous system
Ruflo turns Claude Code from a solo assistant into a self-organizing, cross-machine agent swarm with memory that persists across sessions.

What it does
Ruflo is a plugin-and-CLI harness that wraps Claude Code with multi-agent orchestration. One npx ruflo init installs hooks, an MCP server, and a daemon that routes tasks to specialized agents, maintains vector memory, and even lets agents on different machines collaborate through a federation layer. You keep typing in Claude Code; the swarm handles delegation behind the scenes.
The interesting bit The project is essentially a heavy plugin architecture (33 plugins, 98 agents, 60+ commands) duct-taped onto Claude Code through its extension points. The “self-learning” loop is a feedback arrow in the architecture diagram — agents supposedly learn from successful patterns, though the README is vague on how that learning is stored, evaluated, or prevented from hallucinating confidence.
Key highlights
- Two install paths: lightweight slash-command plugins (zero workspace files) or full CLI install with hooks, daemon, and MCP server
- 33 plugins covering swarms, RAG memory, graph intelligence, test generation, security audit, cost tracking, and even AI crypto trading
- Federation: agents can communicate across machines through a “secure” channel — the README claims no data leakage, but doesn’t detail the threat model
- Memory stack includes vector DB (agentdb), session persistence (rvf), and a knowledge-graph builder
- Also ships a standalone UI beta (flo.ruv.io) and a goal planner (goal.ruv.io)
Caveats
- The “22.2M+ ecosystem downloads” and “115k git clones (14d)” badges link to raw JSON files in the repo, not to npm or GitHub’s own analytics — treat these numbers as unverified
- Windows support exists but the README trails off mid-sentence, suggesting rough edges
- The “Rust-based AI engine” mentioned in the branding is underneath a Cognitum.One architecture; most user-facing code appears to be TypeScript/npm glue
Verdict Worth a look if you’re already deep in Claude Code and need to coordinate many specialized agents across a team or multiple machines. Skip it if you want a simple, well-documented framework — this is a sprawling plugin ecosystem with enterprise surface area and hobby-project documentation depth.