Ruflo

An agent meta-harness — a coordination layer wrapped around claude-code rather than a harness of its own. Formerly Claude Flow, by rUv (ruvnet), MIT, 67.1k stars / 8k forks. Installed with npx ruflo@latest init or as a Claude Code plugin. TypeScript and npm on the surface, Rust underneath, with WASM sandboxes for running agents locally.

What it adds, per the README:

• Swarm coordination — hierarchical and mesh topologies with consensus mechanisms, over 100+ specialized agent roles (coder, tester, security auditor, architect…).
• MCP integration at scale — roughly 210 tools callable in parallel.
• Vector memory — AgentDB with HNSW indexing, so context persists and is searchable across sessions.
• SONA self-learning — agent behavior said to improve from past runs.
• Agent federation — zero-trust collaboration across organizations and machines.
• 12 background workers on auto-trigger, multi-provider routing (Claude, GPT, Gemini, Ollama), a hosted web UI (flo.ruv.io) and a GOAP planner (goal.ruv.io), plus security scanning for CVEs and prompt injection.

Tier: T3 — first-party README, and an unusually expansive one. One number is stated with a condition attached: vector search 1.9×–4.7× faster than brute force above the crossover, which at least names where the comparison stops holding. Nothing else — not the swarm topologies, not SONA, not the 100+ roles — carries a measurement.

Reading it against the corpus

Ruflo is the maximal version of a shape the spoke has several thinner instances of. It stacks, in one package, what the corpus has otherwise seen separately: multi-agent orchestration (orchestrate-100-agents-claude-code does the fan-out with claude -p and nothing else), agent-memory with vector recall and consolidation (iai-pme does that alone, and publishes LongMemEval numbers for it), role personas (auto-company‘s 14, superclaude-framework‘s 20), and sandboxed execution (nono does least-privilege launching alone, from the Sigstore team). Where those projects each defend one claim, Ruflo asserts all of them at once.

That makes the evidence ratio the thing to record. iai-pme is a smaller project that published a retrieval benchmark; jcode publishes self-run performance numbers with versions, ranges and method. Ruflo is an order of magnitude more popular than either and reports one conditional speedup. The star count is the claim doing the most work here, and it measures installs.

SONA is the part that would matter most if it holds: behavior improving from past runs is self-improving-agents in a shipped tool, and the corpus’s existing instance of that idea achieved by filesystem — compound-engineering-plugin‘s /ce-compound, which writes the next run’s context — is deliberately legible about its mechanism. What Ruflo’s learning loop actually updates is not stated.

Federation is the one claim here with no counterpart in the corpus: agents collaborating across organizational boundaries under zero trust is a problem nothing else in the corpus attempts. qm handles the multi-operator case inside one org with a tightening-only security posture; crossing the org boundary is a harder problem and Ruflo is the only entry claiming it.

Related

claude-code · agent-orchestration · agent-memory · iai-pme · superclaude-framework · auto-company · nono · qm · self-improving-agents · model-context-protocol