Three things worth knowing
- Ruflo, a meta-harness built on top of Claude Code and Codex, has 66.3k GitHub stars, with v3.32.22 as its 1,597th release, adding swarm coordination, self-learning memory, and cross-machine agent federation as a layer on top of Anthropic's CLI.
- Two install paths exist with very different footprints: the Claude Code plugin route adds slash commands only, while
npx ruflo initinstalls the full loop with 98 agents, hooks, an MCP server, and a background daemon. - The maintainer dropped a "Flash Attention 2.49x–7.47x speedup" claim from the docs after nightly audits flagged it as unverified, an unusual case of a project deleting its own benchmark number.
Claude Code gives you a capable agent in the terminal. What it doesn't give you is memory that persists across sessions, agents that coordinate without manual orchestration, or a way to share work with agents running on a different machine. Ruflo is built around exactly that gap.
The framing from the ruvnet/ruflo repository is direct: "Agent = Model + Harness. The model writes; the harness gives it tools, memory, loops, sandboxes, and controls so it can actually work." Ruflo is the harness layer. Claude Code stays the execution surface; Ruflo adds what Anthropic doesn't ship.
What Happened
Ruflo shipped v3.32.22, its 1,597th release across 35 contributors. The version bump closed a CLI flag precedence issue for RUFLO_MEMORY_SCAN_ON_WRITE, following a July 26 tracker sweep that closed nine user-reported bugs, including a Windows shell-resolution fix, atomic-write races in daemon state files, and a memory-store dead end on existing keys.
Earlier in the same week, the project rewrote plugin hooks to run natively on Windows, macOS, and Linux via node -e bootstraps instead of /bin/bash -c wrappers that failed on native Windows. The Flash Attention benchmark claim, previously cited in CLAUDE.md as a 2.49x–7.47x speedup, was removed after nightly audits flagged it. The commit is in the repo: the maintainer dropped it rather than kept it.
Key Features
- 100+ specialized agents and swarm coordination: Agents for coding, testing, security, docs, and architecture organize into hierarchical, mesh, and adaptive topologies with Raft, Byzantine, and Gossip consensus protocols.
- HNSW-indexed vector memory (AgentDB): The repo's published audit reports roughly 1.9x faster than brute force at N=20k and 3.2x–4.7x at N=5k, with recall@10 around 0.99. The docs are honest that ANN ties or loses at small N.
- Self-optimizing flywheel (ADR-176): The harness mines its own store into a benchmark, gates candidate configs behind a signed acceptance test (held-out improvement, adversarial pass, drift check, deterministic replay), and promotes winners with a full receipt chain.
- Zero-trust agent federation: Agents on different machines authenticate via mTLS and ed25519, with a 14-type PII detection pipeline stripping data before it leaves your node. HIPAA, SOC2, and GDPR audit trails built in.
- 33 native Claude Code plugins and 21 npm plugins: Spanning RAG memory, security audit, cost tracking, browser testing, and a plugin creator. Individual plugins install with zero workspace footprint.
- Multi-provider routing: Claude, GPT, Gemini, Cohere, and Ollama with failover. The web UI at flo.ruv.io runs the same MCP tools the CLI uses across six frontier models including Qwen 3.6 Max, Claude Sonnet 4.6, and Gemini 2.5 Pro.
Why It Matters
Claude Code's default behavior is session-only. Context, patterns, and decisions disappear when the session ends. Ruflo replaces that with persistent HNSW-indexed memory, 12 auto-triggered background workers, and the repo's claimed 89% task routing accuracy. You install it once with npx ruflo init, use Claude Code normally, and the hooks coordinate in the background.
The federation layer is the part that doesn't exist anywhere else in this space. Agents on different machines or organizations can share work without sharing data. The PII pipeline strips sensitive content before any message leaves your node, and trust upgrades over time based on agent behavior rather than manual configuration.
The benchmark claim deletion is worth naming because it's uncommon. The Flash Attention figure was in the docs, audits caught it was unverified, and it got removed. That's the kind of maintenance discipline that tells you something about how the project is run.
Example Use Case
A TypeScript team wants two organizations to share fraud detection signals without sharing customer data. Team A runs npx claude-flow@latest federation init to generate a keypair, then federation join wss://team-b.example.com:8443. Sending work:
PII is stripped before the message leaves the node, the recipient verifies identity via challenge-response, and both sides get structured audit records searchable via HNSW. federation status shows peer trust levels, which upgrade with history and downgrade instantly on bad behavior with no human intervention required.
For a simpler entry point: install a single plugin like ruflo-rag-memory@ruflo from the Claude Code plugin marketplace to add RAG memory without the full daemon footprint.
Competitive Context
Ruflo doesn't replace Claude Code; it depends on it as the execution surface. The relevant comparison is against LangGraph, AutoGen, and CrewAI as orchestration frameworks. The repo's published benchmark claims wins on cold start, single turn, and RSS by 1.3x to 1953x on darwin-arm64 and linux-x64, with reproduction specs in sota-workload-spec.md. Self-published figures; take the upper end with appropriate skepticism.
The install path split matters for the comparison. The Claude Code plugin route (slash commands, no MCP server) is a meaningfully different product from the full npx ruflo init loop. The plugin path is easy to trial with no workspace impact; the full install is where swarms, hooks, the daemon, and federation actually work. Reading the comparison table in the README before deciding which path to test is worth a few minutes.
My Take
The benchmark deletion is the single most credible signal in this project's recent history. Vendors almost never remove their own numbers. The Flash Attention claim was in the docs, nightly audits caught it, and it's gone. That's the kind of self-policing that I'd want to see more of across the space, and it gives me more confidence in the numbers that stayed.
What I'd still want to see independently: the task routing accuracy claim. The repo says 89%, which is a specific enough figure to be meaningful if it holds across different team setups. It's not in the published audit, and that's the benchmark I'd want to reproduce before committing a team to the full install.
Run your software agents at scale
Cosmos gives your agents the context, tools, and feedback loops they need to get better with every workflow.
Written by
Molisha Shah
GTM
Molisha is an early GTM and Customer Champion at Augment Code, where she focuses on helping developers understand and adopt modern AI coding practices. She writes about clean code principles, agentic development environments, and how teams are restructuring their workflows around AI agents. She holds a degree in Business and Cognitive Science from UC Berkeley.