← Back to research
•·8 min read·project

Claude Flow

Ruflo, formerly Claude Flow, adds coordination and memory around coding agents. See which install path runs, and what the project's own audit rejected.

By Ry Walker, founder of Tembo and Astronomer.

https://github.com/ruvnet/ruflo ↗
Represent this profile? Claim it or send a correction.
profileagent-frameworksopen-sourcerustwasmmcp

Key takeaways

  • The current Ruflo package and the older claude-flow package are both published; the README now leads with ruflo commands.
  • A skill, a lite plugin, a direct MCP connection, and full project initialization provide different capabilities and side effects.
  • The project's own audit rejected headline speed multipliers and documented both working learning mechanisms and subsequent fixes.
  • Current MCP governance is opt-in, so an installed policy file alone does not establish that runtime enforcement is active.

FAQ

What is Ruflo?

Ruflo is the current name of Claude Flow, an open-source coordination and memory layer around coding agents. It supplies a CLI, MCP tools, plugins, and project integration.

Is the package still called claude-flow?

Both ruflo and claude-flow publish version 3.42.0 as of September 15, 2026. Current setup examples use ruflo, while internal packages and configuration still retain claude-flow names.

Does self-learning retrain Claude or another foundation model?

The documented mechanisms reuse memories, update pattern confidence, and adapt routing from feedback. That is different from retraining the commercial model used by the coding agent.

What does Ruflo cost?

The repository is MIT-licensed. Model usage, tools, infrastructure, retries, and human review determine operating cost; more parallel agents do not necessarily reduce the bill.

Executive Summary

Ruflo, formerly Claude Flow, adds orchestration and persistent context around coding agents. The project describes itself as a meta-harness for Claude Code and Codex, with a CLI, MCP server, plugins, hooks, memory, and agent coordination.[1]

Verified September 15, 2026: the latest release is v3.42.0, published September 15 UTC.[2] Both ruflo and claude-flow publish that version; the previous profile's claim that the package name remained exclusively claude-flow is outdated.[3][4]

AttributeValue
CreatorrUv Cohen
LicenseMIT
Runtime requirementNode.js 20 or later, as declared by the published packages
Main surfacesCLI, MCP server, skills, plugins, project initialization
CategoryAgent frameworks

The creator's guide, published package metadata, and license establish these attributes.[5][3][6]

Product Overview

Choose the Installation Surface

The README distinguishes lite plugins from full initialization, while the creator's guide also covers standalone skills and direct MCP connections.[1][5]

SurfaceWhat it suppliesImportant limit
SkillInstructions for working with RufloDoes not itself connect every runtime capability
Claude Code pluginsSelected commands, skills, and agent definitionsCore adds its own MCP server and hooks; other plugins differ
Direct MCP connectionTools exposed by the Ruflo serverDoes not initialize the complete project workflow
CLI initializationProject instructions, agents, hooks, helpers, MCP configuration, and runtime setupWrites workspace configuration and brings a larger operational footprint

The root README calls plugins a lite path without hooks, but the core plugin's own documentation now describes bundled hooks as well as MCP registration. Treat the root table as an imperfect summary, and inspect the actual plugin. Core tools also use a plugin namespace instead of the CLI scaffold's bare tool names.[7][1]

Coordination, Memory, and Optional Extensions

The creator's guide describes task decomposition, role assignment, dependencies, and combining results. AgentDB and RuVector support memory and retrieval; ReasoningBank and related components use prior outcomes and feedback. The guide explicitly distinguishes this from automatically retraining the underlying commercial model.[5]

The optional MetaHarness integration analyzes configuration, declared tool surfaces, and change over time. Its guide documents a degraded mode when optional packages are missing. A successful process exit or available command name is therefore insufficient evidence that a requested audit actually ran; consumers must inspect the result fields.[8]

Federation is a separate cross-machine capability with signed identities, peer trust levels, budgets, and a circuit breaker. Its documentation leaves network connectivity and initial identity vetting to the operator; the WireGuard layer is opt-in.[9]

Technical Architecture

A version-specific initialization example for this review is:

npx ruflo@3.42.0 init wizard

The published ruflo package exposes the ruflo command and depends on @claude-flow/cli. Existing internal names are not evidence that the rename failed or that the older entry point disappeared.[3] The README documents native Windows use through npx, alongside macOS and Linux setup.[1]

Rust and WebAssembly Are Components

The guidance kernel uses Rust compiled to WebAssembly for hashing, secret scanning, destructive-action detection, and relevance scoring. It loads when available and falls back to JavaScript. That supports a narrower description than saying the entire orchestration system moved to Rust.[10]

What the Learning Evidence Establishes

The project's May 29, 2026 audit of CLI 3.10.6 measured persisted pattern-confidence changes and routing feedback. It also rejected several headline multipliers: the Flash Attention speedup was generated rather than measured, and large HNSW claims lacked support in the tested configuration. The same document records removal of randomized telemetry, clearer embedding-backend reporting, and subsequent fixes in 3.10.7–3.10.9.[11]

This is a project-published audit of particular versions and a particular host. It is valuable corrective evidence, not an independent benchmark of 3.42.0. No performance multiplier from that report should be generalized to an end-to-end coding task.

Strengths

  • Several adoption sizes. Skills and selected plugins allow a smaller starting point than full project initialization.[1]
  • Inspectability of optional capabilities. MetaHarness documents availability and degraded-result semantics, which help distinguish configured components from completed work.[8]
  • Published corrective evidence. The internal intelligence audit records both functioning mechanisms and overstated claims, with a remediation history readers can inspect.[11]

Cautions

  • Governance is opt-in. The 3.42.0 release says runtime MCP policy enforcement requires RUFLO_MCP_ENFORCE_POLICY=1; a policy file alone previously had no effect. Verify enforcement through an actual rejected operation before relying on it.[12]
  • Upgrade behavior matters. Direct hive-mind MCP callers now need the token returned by initialization for membership changes and voting. The release says CLI callers handle the token automatically.[12]
  • Hook presence is not a blocking control. The core hook configuration says CLI or installation failures return successfully so they do not block a turn. Treat those hooks separately from MCP governance and host permissions.[13]
  • Cross-machine work adds a trust boundary. Federation's identity and filtering features do not remove the operator's responsibility for peer trust, connectivity, and permissions.[9]
  • Documentation describes different snapshots. The creator's guide pins its initial examples to 3.38.23 and labels later federation additions separately. Confirm installed-version help rather than combining commands from unrelated release examples.[5]

What Developers Say

On May 5, 2026, GitHub user olsenbrands reported a clean-workspace installation study of v3.6.27, while integrating Ruflo into their own multi-agent runtime. They praised idempotent initialization and separate workspace state, but found surprising differences between plugins and CLI setup, incomplete cleanup, and additional context cost on a simple cold request.[14]

This is a useful first-hand report, not a current defect list or a recurring cost estimate. Current core plugin documentation describes MCP registration and hooks that differ from that old report. This refresh did not reproduce the study against 3.42.0.[7]

Pricing & Licensing

Software: MIT-licensed, permitting commercial use and modification subject to the copyright and license notice, without warranty.[6]

Operating costs: Models, tools, infrastructure, retries, and review still cost money or time. The creator recommends measuring cost per accepted result and warns that parallel execution can increase token use even when elapsed time falls.[5]

Competitive Positioning and Fit

Ruflo belongs in the agent frameworks comparison. For readers primarily seeking reusable development instructions, Superpowers and BMAD are also relevant to evaluate; this profile does not establish a performance ranking between them.

Ruflo and Tembo

Tembo is relevant when the decision is how to run coding agents across a team. Its current platform provides cloud or self-hosted execution, shared and resumable sessions, background work from integrations, and centralized visibility. It advertises support for existing agent configurations and repository instructions.[15] Disclosure: Ry Walker is Tembo's CEO and co-founder.

Ruflo exposes coordination, memory, and federation components that an operator assembles and configures. Tembo packages agent execution and team operations as a platform. Evaluate Ruflo when controlling those orchestration components is central; include Tembo when shared environments, handoffs, integration-triggered work, and operational governance are the primary requirements.[5][9][15]

These are overlapping architectural choices rather than a verified integration. A Tembo deployment still needs a review of its supported agents, configuration, permissions, and commercial terms; Ruflo requires the operator to validate the selected components and enabled controls. This review did not run a comparative workload.

Best fit: Teams prepared to evaluate coordination, memory, and optional integrations on a bounded workload, inspect actual tool outputs, and own the runtime configuration.

Poor fit: A small edit that does not justify additional orchestration, or a deployment whose acceptance depends on unverified speedups or security claims.

Viability Assessment

The current release demonstrates continuing maintenance, including corrections to consensus trust weighting, memory-confidence handling, reward feedback, and MCP authorization. It is also evidence that important behavior is still changing. The release's fixes and performance statements remain maintainer claims in this review, not independently reproduced results.[12]

Bottom Line

Ruflo offers a broad set of coordination components, but capability depends on the chosen installation, dependencies, and enabled controls. Evaluate a pinned setup by whether it completes a real task with correct retrieval, visible costs, and tested permissions. Repository activity and feature counts cannot answer that question by themselves.

Research by Ry Walker • methodology

Sources

Share on XShare on LinkedIn