Ruflo gives Claude Code and Codex a 98-agent nervous system with smart routing that sends simple tasks to free tiers automatically. One command to install. 72,566 GitHub stars. 8.1 million npm downloads.
Most AI setups run tasks in a single thread. A request goes in. The model processes it. The result comes out. While that is happening, everything else waits.
That is fine for a chat session. It is a bottleneck for a business running real work through AI.
What Ruflo Is
Ruflo is an agent harness for Claude Code and Codex. Not a different model. Not a replacement for your existing setup. A coordination layer that sits on top of it and adds three things your current setup does not have: parallel agents, persistent memory, and smart routing.
"Agent = Model + Harness," the project documentation says. "The model writes. The harness gives it tools, memory, loops, sandboxes, and controls so it can actually work."
The model you already use stays. What changes is how tasks reach it and what happens after.
How the Parallel Agents Work
In the default single-agent setup, a complex task runs start to finish before the next one starts. Planning happens, then code is written, then tests run, then security gets checked. Four sequential steps.
Ruflo splits the task across specialized agents that run at the same time. One agent handles planning. One writes code. One runs tests. One checks security. All of them share memory across the same run, so the code agent knows what the plan decided and the test agent knows what the code expects.
After each run, the system records which approaches worked. The next run routes faster because Ruflo has already learned the pattern.
The full install gives you 98 specialized agents organized into five categories: core orchestration, memory and knowledge, intelligence and learning, code quality and testing, and security. You do not configure them individually. After running npx ruflo init, you use Claude Code normally and the harness routes in the background.
How the Cost Routing Works
The 75% cost reduction is not a blanket discount. It comes from one specific feature: smart routing.
Not every task you run through Claude Code requires Claude-level reasoning. A lookup, a formatting pass, a simple rewrite: these do not need the full model. They need a fast answer.
Ruflo scores each incoming task and routes it accordingly. Simple tasks go to a free-tier model automatically. Complex tasks that actually need Claude go to Claude. You do not configure this. The harness reads the task and decides.
For a business running AI at volume: client research, content drafts, follow-up emails, report summaries. Most of those are light rewrites or structured lookups. Ruflo stops billing Claude-level rates for tasks a cheaper model handles in the same time.
Two and a half times more output for the same subscription cost is the practical result for a typical mixed workload. If your work is 100% complex reasoning, the savings will be smaller.
2 Ways to Install
There are two install paths. Pick based on how much you want to change.
Path A: Claude Code plugin (5 minutes, no file changes)
Copy this.
/plugin marketplace add ruvnet/ruflo /plugin install ruflo-core@ruflo /plugin install ruflo-swarm@ruflo
This adds slash commands and agent definitions to your existing Claude Code setup. No new files appear in your project directory. Start here if you want to try one agent before committing to the full system.
Path B: Full CLI install (one command, full system)
Copy this.
npx ruflo@latest init
This sets up the complete loop: 98 agents, 60+ commands, 30 skills, MCP server, hooks, and a background daemon. Config files land in .claude/ and .claude-flow/ in your working directory. Use this for production.
For the MCP server alone:
Copy this.
claude mcp add claude-flow -- npx ruflo@latest mcp start
After init, use Claude Code normally. The hooks system routes tasks and coordinates agents in the background without changing how you work.
What Business Owners Use It For
Ruflo ships 35 plugins. A few do things you would otherwise prompt for manually, every run.
ruflo-testgen finds missing tests in your codebase and writes them automatically as part of the same run that writes the code. You do not prompt for tests separately.
ruflo-security-audit scans for vulnerabilities and CVEs during a build run, not after you deploy. Findings come back in the same output as the code.
ruflo-docs generates and updates documentation as code changes. Not on a schedule. As part of the run that made the change.
ruflo-cost-tracker tracks token usage per task type and alerts when a run exceeds budget. Useful if you are running client work through the system and need to know what each project actually cost.
ruflo-federation lets agents on different machines collaborate securely. If you run a team where code, content, and research happen on separate machines, federation routes tasks between them without sending data to a third party.
What to Check Before Relying on It
The full CLI install writes config files, a daemon, and MCP registration to your machine. It is not a lightweight add-on. Run Path A first if you want to see what the agents produce before committing to the install.
The smart routing is good for typical workloads. For anything client-facing where accuracy matters more than speed, review the output the same way you would any AI output. Do not trust auto-routing to handle precision work without a pass.
The 75% cost figure comes from a mixed workload where many tasks are simple. Pure complex reasoning runs will see a smaller reduction.
One Thing to Do This Week
Run npx ruflo@latest init in a project you use regularly. Pick one real task you currently run as a single Claude Code prompt. Run it through Ruflo instead.
Check two things: whether the output quality holds, and what it cost relative to a direct Claude run.
If both look good, route everything through it. If the output needs work on that one task, the fifteen-minute install still told you exactly what to expect at scale.