ruvnet/ruflo

The problem

Claude Code and Codex are single-agent loops: one model, one context window, one machine. When a task grows past that, you end up hand-managing sub-agents, copy-pasting context between sessions, and losing whatever the agent learned last run. Ruflo positions itself as an "agent meta-harness" that sits around those coding agents and adds coordination, memory and routing so multiple agents can work on a task together and carry experience forward.

What's different

Instead of building another agent framework from scratch (LangGraph, CrewAI and similar orchestration libraries), Ruflo plugs into existing hosts — Claude Code via a plugin marketplace, other MCP-enabled clients via a local stdio server — and layers its own runtime on top. The distinguishing pieces are a swarm orchestrator that self-organizes specialized agents, a vector-backed memory layer (RuVector) that feeds past task outcomes back into routing, and federation that lets agents on different machines communicate. That is a different bet from pure libraries: you inherit the host's permissions and tooling rather than replacing it.

Who it's for

It will appeal most to developers already living inside Claude Code or Codex who want parallel task decomposition, cross-session memory, and a visual console without switching editors. Teams experimenting with multi-agent workflows on more than one machine are the federation audience. Anyone who wants a minimal, auditable agent loop, or who is not already committed to Claude Code/Codex, should look at a plain orchestration library instead — the value here is in the host integration, and the surface area is large.

Cost of adoption

Claude Code 2.1.287 or later is required for the plugin path; other clients need MCP configuration, e.g. npx ruflo@latest init wizard. It is MIT licensed and TypeScript/Node-based, so an npm toolchain is a given. The project claims hooks that route and learn automatically, but the README also states mods run with your account's permissions and asks you to review their code — that is real risk for anything touching your repos. The README bundles an enormous amount of marketing (animated badges, a promo for ruOS, an "illustrative download counter"), and describes the learning loop as conceptual, not a live trace; treat capability claims with care until you verify them in your own project.

Verdict

This is an established-looking project by repository age (created June 2025, ~500 days old) with an unusually large star count near 79k, and it just appeared on the weekly board at #4 with roughly 3.3k stars gained. Yet it has never hit the daily board, Hacker News discussion is essentially nonexistent (2 points, 0 comments), and the star-to-attention ratio looks inflated relative to visible community traction. So: real lineage and a real product surface, but the current spike is worth skepticism. Try it in a scratch repo before trusting it with your main workflow.