Rufflow Turns Claude Code Into a 100-Agent Swarm — Free and Open Source

Key Takeaways

• Single install command sets up Rufflow with 100 specialist agents, 60 commands, 30 skills, an MCP server, hooks, and a workspace with pre-configured files.
• Three swarm topologies: hierarchical (manager + specialists), mesh (peer-to-peer coordination where every agent talks to every other), and adaptive (topology switches based on task complexity).
• Vector memory uses HNSW indexing — up to 12,500x faster search than standard memory methods. Memory persists across sessions and is shared across all agents in the swarm.
• Self-learning routing: every task outcome updates a "recipe book" that improves agent selection over time — the system gets better at routing similar tasks to the right specialist.
• Works with Claude Code, OpenAI, Gemini, and Ollama — not locked to a single provider. Swap models without reconfiguring the swarm topology.
• Web UI available at iuv.io for zero-install access: multi-modal chat, 210 tools, persistent memory, and parallel task execution.
• Demo: a 4-agent website team (architect, coder, tester, reviewer) built and deployed a full site in minutes by dividing roles and working in parallel.

How Rufflow Differs from Standard Claude Code

Standard Claude Code gives you one agent and, occasionally, a few sub-agents on complex tasks. Rufflow replaces that with a structured multi-agent system. The analogy used in the video: a restaurant kitchen. The router is the head chef — it reads your order (task), decides which specialist cooks (agents) should handle each part, checks the recipe book (vector memory) for what worked before, and fires them in parallel. The salad agent, the grill agent, and the dessert agent all work simultaneously, then combine their outputs.

The specialist agents cover coding, testing, architecture, research, content, and more — 100 out of the box. When you send a task, the router doesn't just pick one agent; it evaluates which combination of specialists is optimal for the current task, whether multiple agents can work on it simultaneously, and what past experience exists. This last point — the learning loop — means Rufflow's routing genuinely improves the longer you use it.

Vector Memory and the Learning Loop

Most AI tools forget everything when you close the chat. Rufflow stores every task, preference, lesson, and result in a vector memory system using HNSW indexing. The search speed advantage (up to 12,500x over standard methods) matters at scale — when a swarm has processed thousands of tasks, fast retrieval determines whether past experience meaningfully informs current decisions or becomes too slow to query. The memory is shared across all agents in the swarm, so a solution found by one specialist becomes available to all.

The learning loop runs nightly (configurable): it reviews what worked in the day's tasks, updates the routing recipe book with new patterns, and prunes memory entries that proved unhelpful. Over weeks of use, Rufflow's agent selection for your specific workflow becomes substantially better than a fresh install's generic routing.

Swarm Topologies: When to Use Each

Hierarchical: A manager agent oversees specialist workers. Best for structured tasks with clear divisions of labor — building a website, producing a content calendar, running a code review pipeline. The manager handles coordination; specialists handle execution.

Mesh: Every agent communicates directly with every other agent. Best for tasks requiring consensus or shared context — security audits, complex research synthesis, multi-perspective analysis. Slower to coordinate but produces more thorough outputs.

Adaptive: The topology switches between hierarchical and mesh based on task complexity at runtime. Recommended default for general use — it handles both simple and complex tasks without manual topology selection.

Related on OpenClawDatabase

• OpenClaw Hub — Claude Code guides and reference
• OpenClaw Skills Guide
• OpenClaw Configuration
• OpenClaw Cost Optimisation
• Hermes vs OpenClaw — platform comparison