Three prompts, and Claude does the setup
Ruflo is free, open source and genuinely impressive. It is an add-on for Claude Code that lets Claude run a team of agents on one job instead of working through it alone. On 08/09/2026 the repo had 71,400+ stars, 8,461 forks, an MIT licence and a commit from that same day. It is real.
What it is not is free to run. That is the whole reason this page exists, and we get to it below. First, the three prompts.
Install it in a throwaway folder first
The full install writes files into whatever folder you happen to be standing in — a .claude folder, a .claude-flow folder, its own CLAUDE.md, helper files and settings. That is fine in a test folder and annoying in your real project. Prompt 1 handles it for you.
I want to try Ruflo (github.com/ruvnet/ruflo) - a free open-source add-on that lets Claude Code run a team of agents instead of one. Before you touch anything, understand this: I do NOT want it installed into a real project. The CLI install writes .claude/, .claude-flow/, a CLAUDE.md, helper files and settings into whatever folder it runs in. So do it in this order: 1. Make a brand new empty folder called ruflo-test in my home directory and move into it. Confirm out loud that we are inside it before you go any further. 2. Run the interactive setup: npx ruflo@latest init wizard If the wizard asks me questions, show me each one and explain in plain English what the option means before I choose. Do not choose for me on anything that costs money, connects an account, or needs an API key. 3. When it finishes, list every file and folder it created and tell me in one line each what they are for. 4. Show me the exact commands to completely remove it later - delete what it added, and unregister anything it registered. Do not run a swarm or spawn any agents yet. I only want it installed, and I want to know how to undo it.
Ruflo is installed in this folder. I want to see what a swarm actually does on one real job, without setting money on fire. Before you start anything, do these two things: - Tell me how many agents you plan to spawn and why that number. - If this job would be just as good with one agent working through it in order, say so and talk me out of the swarm. I would rather hear that than run it. The job: [REPLACE THIS - pick something with genuinely separate parts. Good example: "go through this project and give me a written report with three sections - security risks, where the tests are missing, and which dependencies are out of date". Bad example: "write me a blog post", because that is one thread of thinking and does not split.] Rules for the run: - Use the smallest number of agents that actually splits the work. Different jobs per agent, never the same job three times over. - Show me which agent has which job before they start. - When they are done, show me each agent's output separately, then the combined answer. - Finish by telling me what this run cost in tokens, and whether one agent doing it in sequence would have been cheaper.
I have been running Ruflo swarms and I want an honest read on whether it is earning its keep. Do not be encouraging. Be blunt. 1. Check whether Ruflo's cost tracker is set up here. Ruflo ships a cost-tracker plugin that tracks token usage, sets budgets and sends cost alerts. If it is not installed, give me the command to add it and tell me exactly what it will show me. 2. Look back at what I have actually used swarms for. For each job, answer honestly: did that need a team, or was it one agent's work with extra steps and extra spend? 3. Set me a token budget I can live with, and show me where the alert fires when I get near it. 4. Then give me two lists for my actual work: three kinds of job where a swarm genuinely wins, and three where it is just a slower, dearer way to get the same answer. If the honest answer is that I should uninstall this and go back to one Claude, say that.
If Claude keeps asking permission, that is normal
Claude Code checks with you before it runs commands on your machine or writes files. When it asks "can I run this?", that is it being polite, not a sign something has broken. Say yes and let it carry on.
What a swarm actually is
Normally Claude is one very capable person. You hand it a job, it works through the job in order, it hands the job back. That is it. One brain, one thread.
A swarm is the same job handed to a small team instead. One agent writes the code, one writes the tests, one reviews it, one documents it — at the same time, sharing a memory they can all read from and write to. When they are done, their answers get pulled back together into one.
Ruflo calls the thing that runs the team a Queen. The Queen decides who does what and in what order, and when the agents disagree, they vote. The README names three voting systems — Raft, Byzantine and Gossip. You will never pick one by hand. It matters only because it tells you this is an actual coordination system underneath, not three chat windows open side by side.
100+ specialist agents
Coder, tester, reviewer, architect, security and more. The capability table says 100+; the install-path table says 98. Either way, not 60.
Queen-led coordination
Hierarchical, mesh and adaptive team shapes, with consensus when agents disagree. You describe the job; it decides the shape.
Memory that survives
Vector memory (AgentDB with HNSW indexing) so the team remembers your project between sessions, not just inside one chat.
It learns from your runs
SONA patterns, ReasoningBank and trajectory learning — meant to get better at routing your work the more you use it. They claim 89% routing accuracy.
12 background workers
Auto-triggered jobs that fire without you asking — auditing, optimising, hunting for gaps in your tests.
Five providers, one router
Claude, GPT, Gemini, Cohere and Ollama, with failover. Plus a plugin marketplace and cross-machine agent federation if you ever get that far.
There are two very different installs, and the reels do not say which
The plugin path adds slash commands only. Zero files in your workspace. Run /plugin marketplace add ruvnet/ruflo then /plugin install ruflo-core@ruflo. Good for a look around.
The CLI path is the one people are demoing. npx ruflo@latest init wizard gives you the whole loop — the README's own table says 98 agents, 60+ commands, 30 skills, an MCP server, hooks and a background daemon — and it writes into your folder. This is the one worth putting in a test directory first.
What the reels got wrong
I say this on camera and I want the receipts here in writing. Ruflo deserves the attention it is getting. The numbers attached to it do not survive five minutes with the actual README, and I read the whole thing on 08/09/2026.
What is going round
- It runs 60 agents
- It halves your token use
- It stretches your Claude usage by 250%
What the README says
- 100+ agents, not 60. The install-path table says 98. Nobody anywhere says 60.
- No token-reduction claim exists. Search the file for "50%" — it is not there. There is no efficiency or savings claim of any kind.
- "250%" does not appear either. Not in the feature list, not in the benchmarks, not in the docs index.
What IS in there points the other way
Ruflo ships a plugin called ruflo-cost-tracker, described in its own feature list as: track token usage, set budgets, get cost alerts. Its live agent dashboard shows each spawned agent's token budget so you can kill runaway workers. You do not build budget alarms and a kill switch for a tool that saves you money. You build them for a tool that can spend fast.
How I checked, so you can too
Repo: ruvnet/ruflo on GitHub. On 08/09/2026 it showed 71,400+ stars, 8,461 forks, an MIT licence, a commit that day and no archive notice — so this is a live, healthy project, not a dead one. I then searched the full README for the three figures. The only percentages in the entire file are an 89% task-routing accuracy claim and some digits buried inside an image link. No 50%. No 250%. No 60.
The speed numbers it does publish are about memory lookups and cold-start time, not about how many tokens a job costs you. Different thing entirely.
When a hundred agents beats one Claude
The honest test is not "is this impressive". It is "does this job actually split?" If the parts of the work need each other's answers, a team cannot help you — they just wait around and bill you for waiting.
The one-sentence rule
If you cannot say what the second agent's job is in one sentence — a different job to the first one — you do not need a second agent. Every time I have wanted a swarm and could not pass that test, one Claude did it better and cheaper.
The honest catch
Here is the bit I want in writing, because it is the exact opposite of what the viral version implies.
A hundred agents on a small job burns your limit faster, not slower
Every agent is its own conversation. Its own context, its own instructions, its own back-and-forth with the model. Ten agents on one job is not one job's worth of usage divided ten ways — it is closer to ten jobs' worth, plus all the coordination chatter between them.
A swarm buys you two things: work happening in parallel, and separate opinions you can compare. It does not buy you a discount. Nothing about running more agents makes each one cheaper.
None of which means don't use it. It means use it with your eyes open. Four things that keep it sensible:
Start in a throwaway folder
The CLI install writes into whatever directory it runs in. Look at what it added before you point it at anything you care about.
Install the cost tracker on day one
Not after your first surprise. Budgets and alerts are cheap insurance, and Ruflo already ships them — that is a hint from the people who built it.
Cap the agent count yourself
The instinct is more. The useful number is usually three or four, doing genuinely different jobs. A hundred available does not mean a hundred at once.
Make it justify the swarm
Prompt 2 above tells Claude to talk you out of it when one agent would do. Leave that line in. It has saved me more than it has cost me.
And the part that is not a catch
It is MIT-licensed and free. It had a commit on the day I checked it. If your work genuinely splits into separate parts, this is a serious piece of kit and you should go and play with it. Just go in knowing what it costs to run, rather than believing it pays for itself.
Ready to go deeper?
Swarms are the flashy half. The useful half is knowing which of your jobs actually splits.
Wright Mode Membership
Join a community of women using AI to take real work off their plate. Live calls, copy-paste prompts and workflows, and ongoing support.
Claude Masterclass
Learn to use Claude like a pro — from prompting fundamentals to building real workflows that save hours every week.
Claude Code Masterclass
Go beyond the chat box. Build automations, process data and create tools with Claude Code — no developer experience needed.