Ruflo, formerly Claude Flow, positions itself with a formula the README states twice: “Agent = Model + Harness.” and “Ruflo is the harness.” It is an execution layer wrapped around Claude Code and Codex, adding specialised agents, swarms, memory and federation. MIT licensed, held by RuvNet.
The feature list is expansive: coordination of “100+ specialized agents” in swarms with “Hierarchical, mesh, and adaptive topologies with consensus”, vector memory via AgentDB with HNSW indexing, “SONA neural patterns, ReasoningBank, trajectory learning”, background workers, cross-machine federation with mTLS and ed25519 identity verification, a plugin marketplace and a console dashboard.
The README’s own caveats are the most useful thing in it
Projects at this level of ambition usually oversell. Ruflo’s documentation does something more unusual: it marks its own limits, and those marks are what you should read first.
- The architecture diagram — “User → Ruflo (CLI/MCP) → Router → Swarm → Agents → Memory → LLM providers” — is labelled “a conceptual flow, not a live execution trace.”
- The console animations use “sample data,” not a live session.
- Benchmark claims come from an external gist and are the project’s own measurements.
- The HNSW retrieval speedups apply only above a crossover point: it “ties/loses at small N.”
- Learning results “depend on the configured memory, learning components and quality of feedback,” and local vector retrieval still requires “embeddings and configured learning modules.”
- Hook support “depends on the client.”
Taken together those tell you what to expect: a substantial harness whose more advanced claims — neural patterns, trajectory learning, swarm consensus — are capabilities you must configure and feed, not behaviour you get by installing. The README does not pretend otherwise, which is to its credit, but the distance between the feature list and a working setup is real.
One inconsistency is worth noting as a maintenance signal rather than a flaw: the README cites “33 native Claude Code plugins + 21 npm plugins” in one place and “All 35 plugins” in another.
The security line to read before installing
This one is not a caveat about accuracy — it is about what runs on your machine. The documentation states that mods “run with your account’s permissions and are not sandboxed, so read the code first.”
That is the correct warning and it should change your behaviour. An orchestration layer that installs plugins which execute unsandboxed with your credentials is a meaningful trust decision, not a package install. The console’s Claude control sensibly defaults to “read + ask”; raising permission levels is explicitly the user’s choice, and it is a choice worth making slowly.
Installing
The plugin path requires Claude Code 2.1.287 or later:
/plugin marketplace add ruvnet/ruflo
/plugin install ruflo-core@ruflo
/reload-plugins
/rufloOr via the CLI:
npx ruflo@latest init wizard
npm install -g ruflo@latestAs an MCP server for Claude Code:
claude mcp add claude-flow -- npx ruflo@latest mcp startOther MCP clients use command npx with arguments ["-y", "ruflo@latest", "mcp", "start"]. There is a curl … | bash script for macOS, Linux, WSL and Git Bash; the documentation notes it does not run in native Windows PowerShell or cmd.
Who this is for
Ruflo is for someone who has already outgrown a single agent session and wants coordination, shared memory and multi-machine work — and who has the appetite to configure the memory and learning components rather than expecting them to arrive working.
If that is not you, the cost is high. A hundred agents, swarm topologies and unsandboxed plugins are a lot of surface to take on for a workflow that one well-instructed agent handles.
The reasonable way in is the MCP server plus the console at its default “read + ask” permission, on a repository you do not mind an agent exploring — then decide whether the swarm features earn the configuration work. And take the project at its word on reading plugin code first.
