Claude Flow

Overview

Claude Flow — renamed Ruflo in early 2026 — is an open-source agent orchestration platform that transforms Claude Code into a multi-agent development system. The project now describes itself as "the leading agent meta-harness for Claude," coordinating 100+ specialized agents — coders, testers, reviewers, architects, security auditors — working together on complex software engineering tasks.^[1]

Status (verified June 11, 2026): Very actively developed — the latest release (v3.10.42) shipped the morning of this verification, capping a near-daily release cadence through May and June 2026.^[2] The repo sits at 58,989 stars and 6,782 forks, up from ~14k stars when this profile was first written in February 2026.^[1] Rename: the repository moved from ruvnet/claude-flow to ruvnet/ruflo with the v3.5 stable release (Feb 27, 2026), reportedly for trademark reasons; old GitHub URLs 301-redirect to the new location, and the npm package and CLI command remain claude-flow (v3.10.42, MIT, not deprecated).^[3][4]

The v3.5 Rewrite: Rust/WASM Kernel

The rename coincided with a genuine architectural shift, not just a rebrand. With v3.5 (stable Feb 27, 2026), the system's hot path moved from Node/TypeScript to Rust compiled to WebAssembly: the policy engine, embeddings, and proof system now run as WASM kernels, offloading operations like hashing, secret scanning, destructive-action detection, and shard scoring from JavaScript.^[3][5] The project claims up to 352x faster execution for certain operations via the WASM layer — like most Ruflo performance numbers, this is self-reported.^[3]

The name itself nods to the shift: "Ruflo" is a portmanteau of Ru (Rust, and creator Reuven/rUv) and flo (flow states).^[3]

Architecture

The system follows a layered architecture:^[1]

1. Entry Layer — CLI and MCP server with AI-based security (AIDefence)
2. Routing Layer — Q-Learning router with Mixture of Experts, skills, and hooks
3. Swarm Coordination — Configurable topologies (mesh, hierarchical, ring, star) with consensus protocols (Raft, BFT, Gossip, CRDT) and human-agent coordination via Claims
4. Agent Layer — 100+ specialized agents dispatched based on task type (up from 60+ pre-v3.5)^[3]
5. Resource Layer — Memory (AgentDB), multi-provider LLM support, and background workers

RuVector Intelligence Layer

The self-learning subsystem includes:

• SONA — Self-optimization with sub-0.05ms claimed latency
• EWC++ — Elastic weight consolidation to prevent catastrophic forgetting
• Flash Attention — claimed 2.49–7.47x speedup on attention operations
• HNSW — Vector search with claimed 150x–12,500x faster retrieval
• ReasoningBank — Pattern storage for reuse across tasks
• 9 RL algorithms — Q-Learning, SARSA, PPO, DQN, and more for continuous routing improvement

These figures come from the project's own documentation; independent benchmarks are scarce. Notably, the project's own v3.10.30 release notes report its retrieval stack ranking 3rd of 11 on a 4-dataset BEIR evaluation — a candid mid-pack result that contrasts with the headline multipliers.^[2]

Key Differentiators

• Swarm-native: Not just multi-agent — swarm coordination with consensus and topology selection
• Rust/WASM core: Post-v3.5, hot-path operations run in compiled Rust kernels rather than Node^[5]
• Self-learning loop: RETRIEVE → JUDGE → DISTILL → CONSOLIDATE → ROUTE cycle continuously improves agent routing
• MCP-first: Built as an MCP server, composable with Claude Code, Codex, and the broader MCP ecosystem^[1]
• Multi-provider: Works with Claude, GPT, Gemini, and Ollama — not locked to Anthropic

Cautions

• Self-reported benchmarks. The signature performance claims (150x–12,500x retrieval, 352x WASM speedups, sub-0.05ms SONA latency) are the project's own numbers. No independent third-party benchmarks were found as of June 2026.
• Complexity vs. need. Consensus protocols, RL routing, and vector intelligence layers are heavy machinery for most coding tasks. Even sympathetic reviewers flag this (see below).
• Naming churn. The repo is ruflo, the npm package and CLI are claude-flow, and docs mix both names — confusing for newcomers, and a long tail of tutorials still point at the old URLs.
• Bus factor. Despite ~59k stars, the project remains dominated by a single prolific author; the near-daily release cadence cuts both ways for teams wanting stability.
• Hype-to-evidence gap. Community discussion of Claude Flow has long mixed enthusiasm with skepticism about whether the elaborate architecture delivers proportionate real-world results; production case studies remain anecdotal.

What Developers Say

Independent, attributable reviews remain thin relative to the project's star count as of June 11, 2026 — most coverage is either the author's own material or unattributed social posts. The most substantive third-party assessments come from engineer Steven Gonsalvez's two-part analysis:

"...the future of how we build software or the most over-engineered tool in human history. Possibly both." — Steven Gonsalvez, on Ruflo's swarm architecture^[3]

"If you need one agent to fix a bug, it's massive overkill." — Steven Gonsalvez^[6]

Gonsalvez also credits the v3.5 rewrite as substance over marketing: the move to Rust/WASM kernels "represents serious engineering commitment to performance."^[3] No substantive Hacker News or Reddit threads with quotable firsthand production reports were found during this verification — itself a notable signal for a ~59k-star project.

Pricing & Licensing

Licensing model: MIT — free for commercial and personal use.^[1]

Hidden costs: You pay for underlying model usage (Anthropic/OpenAI/Google API keys or a Claude Code subscription); local Ollama is the zero-API-cost path.

Competitive Context

Ruflo occupies the "swarm orchestration" niche — heavier than simple multi-agent frameworks like CrewAI, more research-oriented than production platforms like Tembo. The climb from ~14k to ~59k stars between February and June 2026 signals exceptional community interest in autonomous agent swarms,^[1] though the complexity of the system (consensus protocols, RL routing, vector intelligence layers) still raises the question of real-world production adoption vs. experimental usage — a question the thin base of independent user reports has yet to answer.

Bottom Line

Ruflo (née Claude Flow) is the most ambitious open-source agent orchestration project in the Claude ecosystem, and the v3.5 Rust/WASM rewrite shows the engineering is real, not just branding.^[3] Star growth and release velocity are exceptional; independent evidence that the swarm machinery outperforms simpler harnesses is still missing.

Recommended for: Experimenters and researchers exploring swarm-based agent coordination, and teams already deep in Claude Code/MCP who want a maximalist harness to study or extend.

Not recommended for: Teams that need one agent to fix bugs reliably — simpler harnesses do that with far less machinery — or anyone who requires independently benchmarked, production-validated tooling.

Outlook: The rename decouples the project from Anthropic's trademark and the Rust kernel positions it for multi-model "meta-harness" ambitions (Codex integration is already first-class). Watch for independent benchmarks and named production users; until those appear, treat the headline performance multipliers as vendor claims.

---

Research by Ry Walker Research • methodology