"60 Claude agents in parallel and 75% off your API bill" — that's how RuFlow gets sold on social media. The truth is right next door — and it's more interesting than the hype 🐝
What it is and who it's for
A single Claude Code chews through tasks one at a time: planned, coded, tested — all by itself, all sequentially. RuFlow (officially Ruflo; until early 2026 the project was called Claude Flow) is an open-source "meta-wrapper" by Reuven Cohen (ruvnet) that turns a lone agent into a coordinated swarm: one agent plans, a second codes, a third tests, a fourth checks security — in parallel and with shared memory. ★64k+ on GitHub, MIT license, fresh releases almost daily (v3.28 as of July 2026).
Who for: people who already live in Claude Code, have hit the ceiling of a single agent and are ready to dig into a serious system — this is not install-and-forget.
Key features
- Agent swarm: a full install gives you 98 specialized agents (coder, tester, reviewer, architect, security engineer…), 60+ commands, 30 skills, an MCP server, hooks and a background daemon; the repo's marketing says "100+ agents".
- Smart model routing: tasks are spread across 5 providers (Claude, GPT, Gemini, Cohere, Ollama) with failover; the README claims 89% accuracy for the intelligent router, and the ruflo-ruvllm plugin runs simple steps on local models.
- Shared vector memory (AgentDB + HNSW): agents remember context between sessions and self-train on successful patterns (SONA).
- Swarm topologies: hierarchy with a "queen", mesh and adaptive; Raft / Byzantine / Gossip consensus.
- 35 plugins: RAG memory, test autogeneration, security audit, observability, a token spend tracker (ruflo-cost-tracker) and more.
- Federation: agents on different machines collaborate over zero-trust (mTLS + ed25519), and PII is stripped from messages automatically.
- Web UI (beta) at flo.ruv.io: a multi-model chat with parallel MCP calls — you can try it without installing anything, no account, no key.
Getting started in 5 minutes
- You need Node.js and Claude Code. The full route:
npx ruflo@latest init wizard— an interactive wizard that works in both PowerShell and bash. - The light route with no files in your project: in Claude Code run
/plugin marketplace add ruvnet/ruflo, then/plugin install ruflo-core@ruflo— slash commands only, no MCP server and no hooks. - The MCP server connects to Claude Code with a single command:
claude mcp add ruflo -- npx ruflo@latest mcp start. After that just work as usual — the hooks route tasks and spin up agents in the background on their own.
When to take it and when not to
✅ Take it if Claude Code is your main tool and you need parallel roles on one task: plan, code, tests, security review with shared memory. ✅ Take it if you want memory between sessions and want simple steps offloaded to cheap/local models (Ollama) — it's all open source, no subscriptions. ❌ Skip it → plain "Claude Code" is enough: for single tasks the built-in subagents do the job, and a swarm is overkill. ❌ Skip it → go with "Dify" or "Flowise" if you need a visual LLM workflow builder without a terminal. ❌ Skip it → go with "Aider" if you need a lightweight coding agent in the terminal without an ecosystem of 35 plugins.
The honest price
100% free and open source (MIT), no subscriptions — tagged free / oss / mcp. But the tokens come out of your Claude plan or API key: a swarm of dozens of agents burns them faster than a single agent. The viral "75% off your API spend" is a social-media number: it's nowhere in the official README. The savings are real mechanically (simple steps go to cheap and local models instead of the top one), but the exact percentage depends on your workload — measure it yourself with the ruflo-cost-tracker plugin. The flo.ruv.io web demo is free and needs no key.
Gotchas
- Name confusion: the project started as Claude Flow, is now officially Ruflo, and shows up on social media as RuFlow/RuFlo. Search the docs for "ruvnet ruflo".
- Viral posts credit the project with numbers ("75% off API", old SWE-Bench benchmarks) that aren't in the current README — check against the repository.
- Two install routes with different functionality: the Claude Code plugin is the light version with no MCP server and no hooks; the full swarm only comes from
npx ruflo init. They're easy to mix up. - A full CLI install writes a pile of files into your project:
.claude/,.claude-flow/,CLAUDE.md, helpers and settings. - This is enterprise scale: for small tasks the community flat-out calls it overkill — the learning curve is noticeable.
- A parallel swarm burns tokens in parallel: without budgets and limits the bill can be an unpleasant surprise.
- The project moves fast (repo created in June 2025, already v3.28 by July 2026): commands and APIs change, and six-month-old tutorials are already wrong.
🤖 Prompt accelerator
"I've installed Ruflo (npx ruflo@latest init) on top of Claude Code. Take the task
and break it out across the swarm: the planner agent writes the plan, the coder implements it, the tester generates tests, the reviewer checks security. Route simple steps to the cheap model and complex ones to the main one. Save key decisions to memory (memory_store) so the next run doesn't start from scratch. At the end, show which agent did what and how many tokens it took."