Ruflo

Ruflo sits on top of Claude Code, Codex or Hermes and turns one agent into a swarm. Install it as an MCP server, and your host agent gains planners, testers, reviewers and a shared memory it can query. The trait that decides it is appetite: it is for the power user who enjoys tuning a system and has a backlog that justifies a hundred agents.

Pick it if you want the maximal version of multi-agent and will read the user guide. Pick oh-my-claudecode if you want a staged pipeline with fewer moving parts, and Claude Code alone if one agent is finishing your work.

The risk is context fan-out. Roughly 210 MCP tools across five server groups are exposed to the host agent, and every tool schema costs tokens on every turn before any work starts. The quick-start command is claude mcp add claude-flow -- npx ruflo@latest mcp start, so the old name is still the identifier a user will see, and mixed names in a config are how people install two copies.

Load one server group at a time. What it does right: it never touches files itself; the host agent performs every edit and command, so a swarm cannot write what the host would refuse.

The documentation is dense with mechanism. 1. Swarms run in hierarchical, mesh or adaptive topologies, with a queen-led hierarchy that names Raft, Byzantine and Gossip consensus. 2. Memory is an HNSW-indexed vector store; the README reports recall at 10 of about 0.99 and a 1.9 to 4.7 times speed-up over brute force. 3. Benchmarks against LangGraph, AutoGen and CrewAI claim wins from 1.3 to 1953 times on cold start and memory.

Those numbers live on a perf branch with reproduction notes, better than most, but the 1953 figure is a ratio of startup costs, not of task outcomes. Consensus among language-model agents is a metaphor until a paper shows otherwise.

rUv is a person, not an entity with a cap table. The project has 70,339 stars, an MIT licence and two hosted surfaces, a multi-model chat at flo.ruv.io and a goal planner at goal.ruv.io, neither of which has a published price, so the hosted layer is a demo until it is not. Pricing power: none yet, and the audience is people who will not pay.

Moat: the author's velocity, which is also the bus factor. Likely exit: acqui-hire by a lab or a dev-tools company that wants the swarm vocabulary; likely pivot: a paid hosted planner. Position: watch, do not depend.

The demo is a swarm finishing a feature. Procurement: there is no vendor, no seat and no contract, so the cost is entirely the meter, and a tool whose selling point is a hundred agents multiplies whatever each engineer already spends. Twelve background workers trigger automatically, which is spend nobody scheduled. No SSO, no central log, no retention policy beyond the host agent's.

Onboarding is npx ruflo@latest init wizard, then a user guide longer than most of our runbooks. It does not run in CI. Not yet; revisit if a hosted tier with spend caps and an admin appears.

MIT and all of it on GitHub. npx ruflo@latest mcp start brings up the MCP server, the model router takes Ollama and OpenRouter alongside the hosted labs, so a swarm can run on my own box, and the 27-hook system lets me intercept routing before it happens. The Claude Code plugin path is a marketplace add, and nothing here needs a login.

The grudging part is scale: there is more surface here than I will ever configure, and reading it is a project. But everything is a file and everything is forkable. Respect, with a raised eyebrow.