ruvnet/ruflo

🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated

Ruflo – an agent meta‑harness for Claude Code & Codex

What it is – Ruf Flo is a harness that sits around large‑language‑model‑based coding assistants (Claude Code, Codex) and turns a single model into a full‑featured, self‑organising AI team. It provides a CLI, a plugin marketplace, and a web UI that expose 100 + specialised agents, persistent vector memory, swarm coordination, federation across machines, and security guardrails.

How it works – The core loop looks like:

User → Ruflo CLI / MCP server → Router → Swarm → Agents → Memory → LLM providers
                         ^                               |
                         +---- Learning loop <------------+

The harness adds tools, loops, sandboxing and a learning layer so that agents can plan, remember, test, improve and collaborate without the user having to manage any of those details.

Key capabilities (as listed in the README)

Capability What you get
100 + agents Ready‑made agents for coding, testing, security, documentation, architecture, trading, IoT, etc.
Swarm coordination Hierarchical/mesh topologies, consensus, adaptive routing.
Self‑learning loop Agents record successes, update their behaviour (SONA patterns, ReasoningBank).
Vector memory (AgentDB) HNSW‑indexed memory, ~0.99 recall@10, 1.9‑4.7× faster than brute‑force for 5‑20 k entries.
Federation Zero‑trust, authenticated communication between agents on different machines or organisations.
Multi‑provider routing Claude, GPT, Gemini, Cohere, Ollama, custom OpenAI‑compatible endpoints.
Security & compliance Prompt‑injection guard, PII detection, CVE scanning, input validation.
Meta‑harness audit Tool to grade an agent setup (1‑100), snapshot configs, detect regressions.
Web UI (beta) Multi‑model chat with parallel MCP tool calling, persistent memory, Docker‑ready self‑hostable deployment.
Goal Planner UI GOAP‑style plain‑English goal decomposition → A* planning → live agent execution.
Plugin marketplace 33 Claude Code plugins + 21 npm plugins covering core, memory, intelligence, DevOps, domain‑specific use‑cases.

Installation paths

Path What you install When to use
Claude Code plugin (lite) Only slash‑commands and a handful of agent definitions. No files added to your repo. Quick try‑out of a single plugin without committing to the full stack.
Full CLI install (npx ruflo init) All 98 agents, 60+ commands, MCP server, hooks, daemon, .claude/ & .claude‑flow/ config files. Production use where you want the complete orchestration layer.

CLI quick‑start (macOS/Linux/WSL/PowerShell):

# one‑liner (POSIX shells)
curl -fsSL https://cdn.jsdelivr.net/gh/ruvnet/ruflo@main/scripts/install.sh | bash
# or, cross‑platform
npx ruflo@latest init wizard   # interactive wizard
# non‑interactive
npx ruflo@latest init

You can also install globally with npm i -g ruflo.

MCP server registration (for Claude Code):

claude mcp add claude-flow -- npx ruflo@latest mcp start

Ecosystem & plugins (selected categories)

  • Core & orchestration – ruflo-core, ruflo-swarm, ruflo-autopilot, ruflo-federation.
  • Memory & knowledge – ruflo-agentdb, ruflo-rag-memory, ruflo-ruvector (GPU‑accelerated vector search), ruflo-knowledge-graph.
  • Intelligence & learning – ruflo-intelligence, ruflic‑graph-intelligence, ruflo-ruvllm (local LLM routing), ruflo-goals.
  • Code quality – ruflo-testgen, ruflo-browser (Playwright), ruflo-jujutsu (git‑diff risk scoring), ruflo-docs.
  • Security – ruflo-security-audit, ruflo-aidefence.
  • Methodology – ruflo-adr, ruflo-ddd, ruflo-sparc, ruflo-metaharness, ruflo-arena.
  • DevOps – ruflo-migrations, ruflo-observability, ruflo-cost-tracker.
  • Domain‑specific – ruflo-iot-cognitum, ruflo-neural-trader, ruflo-market-data.

Typical use‑cases

  1. AI‑assisted software development – Write code, automatically generate missing tests, run browser‑based integration tests, and get security audits—all orchestrated by agents.
  2. Goal‑driven automation – Declare a high‑level objective (e.g., "launch a new micro‑service"); the Goal Planner decomposes it, assigns tasks to agents, and tracks progress.
  3. Enterprise‑scale agent federation – Multiple teams run Ruflo instances; agents securely exchange work without leaking data.
  4. Self‑hosted AI stack – Run the web UI and the ruvLLM local model stack behind a firewall for fully offline development.
  5. Financial‑tech or IoT operations – Use the domain‑specific plugins (neural-trader, iot‑cognitum) to build autonomous trading bots or fleet‑management agents.

License & community

  • License: MIT (per badge).
  • Stars / downloads: 8.1 M+ ecosystem downloads, ~106 k Git clones in the last 14 days (as reported in the README).
  • Documentation: 14‑chapter “Ruflo Explained” guide, ADRs, benchmark scripts, and a plugin‑creator scaffold.

Bottom line

Ruflo is a full‑stack, plug‑in‑driven meta‑harness that turns Claude Code / Codex into a collaborative AI team with memory, learning, security and cross‑machine federation. It can be tried instantly via the hosted UI or installed locally with a single npx ruflo init command, making it a concrete, production‑ready platform for agentic software engineering and beyond.

Related

  • Project
    vibeeval/vibecosystem

    vibecosystem is an open‑source package that expands Claude Code into a self‑organising AI development team. It ships 138 specialised agents, 296 reusable skills, 74 TypeScript hooks and 20 behavioural rules. After a one‑line install you can ask Claude Code to “add a feature” and the swarm of agents will plan, code, review, test and learn automatically. Profiles let you enable only the agents you need to save tokens, and a `vibeco` CLI plus a web dashboard give stats, health checks and token‑cost tracking. All data stays locally in `~/.claude/`; no telemetry is sent. The project also supports Cursor, Codex CLI and OpenCode via dedicated install scripts.

  • Project
    SuperClaude-Org/SuperClaude_Framework

    SuperClaude Framework is an open‑source meta‑programming platform that extends Claude Code with 30 slash commands, 20 AI agents, 7 behavioral modes, and optional MCP servers for faster, token‑efficient execution. It automates the full software development lifecycle—from brainstorming and design to coding, testing, documentation, and project management—via simple commands like `/brainstorm`, `/implement`, and `/research`. Installable via pipx (or from source) and licensed under MIT.

  • Project
    borghei/Claude-Skills

    Claude Skills is an open‑source library of 385 ready‑to‑use “skills” – prompt‑plus‑Python bundles – that can be loaded into Claude, ChatGPT, Gemini, Copilot, Cursor and other LLM assistants. It ships 77 expert personas, 900 pure‑Python tools, and compliance workflows for 18 regulated frameworks, letting non‑engineers run real business, product‑management and compliance tasks via slash‑commands or plugin installation.

  • Project
    shinpr/claude-code-workflows

    A set of structured development workflows and plugins for Claude Code that ensure AI-generated code stays within scope and is verified against agreed-upon design documents.

  • Project
    Ancienttwo/repo-harness

    repo‑harness is a Bun‑based CLI that makes Claude or Codex agents work against a Git repo by storing all session state (plans, contracts, handoffs, checks, reviews) in files. Hooks enforce that each LLM call respects these artifacts, enabling deterministic resumption, token‑efficient context, and auditable evidence. An optional “authorized program” layer lets autonomous controllers run unattended work under operator‑minted authorizations, budget caps, and renewable leases.