Ruflo is a free, open-source add-on that lets Claude Code run many specialised agents together, with shared memory between runs. This guide covers what it is, the two ways to install it, and what to treat with care before you point it at real work.
What you need
Claude Code, installed and signed in. Claude Code needs a paid Claude plan.
Node.js and npm on your computer. One install guide lists Node 18 or newer.
A test project in Git, so you can undo anything Ruflo changes.
Ruflo itself: ruvnet/ruflo on GitHub. MIT licence, free. The repo showed about 77.9k stars in October 2026.
Step by step
Know what you're installing. Ruflo is built by ruvnet and was originally called Claude Flow. It works as a "meta-harness" around Claude Code and Codex:
Agents: 100+ specialised agent types, such as coder, tester, reviewer, architect and security. The full install sets up 98 of them. Coordinator agents split a task and hand parts to workers that run in parallel.
Shared memory: a local database (AgentDB, with HNSW vector search) that stores decisions and files so later runs can recall them.
Routing: simple code edits go to "Agent Booster", which applies them with local WebAssembly and no model call. Medium tasks go to Haiku or Sonnet. Complex tasks go to Opus. These tiers are Ruflo's own design, as described in its docs.
Pick your route. The plugin route adds slash commands only. It's light and easy to remove. The full
initroute installs 98 agents, hooks, an MCP server and a background daemon. Start with the plugin.Plugin install (inside Claude Code):
/plugin marketplace add ruvnet/ruflo
/plugin install ruflo-core@ruflo
/plugin install ruflo-console@ruflo
/reload-plugins
/rufloFull install (in your terminal, inside the test project):
git add -A && git commit -m "before ruflo"
npx ruflo@latest init wizardTo use Ruflo's tools through MCP in Claude Code:
claude mcp add claude-flow -- npx ruflo@latest mcp startRun a first task. Give it something small and well defined, and ask for a plan before any code:
Use Ruflo's agents for this task: add input validation and unit tests to the signup form in this project. First show me the plan: which agents you'll use (planner, coder, tester, security reviewer), what each will do, and which model tier each step needs. Wait for my OK before changing files.Review and compare. Check the diff with
git diff. Then do the same task in plain Claude Code and compare your usage. That tells you whether the swarm actually saves you anything on your kind of work.
Good to know
Cost claims are self-reported. You'll see "75% lower API costs" in some guides and "250% more subscription capacity" in others. Ruflo's own docs elsewhere claim 30 to 50% token savings. None of these has been independently verified, and the current README doesn't repeat the 75% figure.
There is no "free tier" model. The "free" part is Agent Booster handling simple code edits locally, with no model call. Everything else still uses your Claude usage.
More agents can mean more tokens. For comparison, Anthropic's own Agent Teams feature uses roughly 4 to 7 times the tokens of a single session, according to one analysis.
Not "number one". The README itself says its lineage claim is not "a universal first across every Claude Code tool". Its benchmark wins over LangGraph, AutoGen and CrewAI are self-published, and one earlier speed claim was removed after an audit.
Permissions. Some older guides start Claude Code with
--dangerously-skip-permissions. Don't. Keep approvals on, especially with a background daemon running.Use a test repo first. The full install adds hooks and config to your project. Commit before you install so you can roll back.
Sources
Found this useful? Follow @kevinsingh.me on Instagram for a new AI tool every day, and subscribe so you don't miss the next one.