RooFlow v3.6.12: Federated Agent Swarms with 314 Tools for Claude Code

Key Takeaways

• RooFlow is the new name for ClaudeFlow, renamed January 2026 — 36,000+ GitHub stars, ~500,000 downloads, ~100,000 monthly active users across 80+ countries.
• 60+ built-in agent roles (coder, tester, reviewer, architect, security, writer, researcher, documenter); all share one memory via AgentDB, which is 150–12,500x faster than basic search.
• AgentDB memory is persistent across sessions, projects, and machines — you can shut your laptop, come back the next day, and agents pick up where they left off.
• Federation (new in v3.6.12): two RooFlow setups on different machines can share agents and work together using mTLS + ED25519 key authentication — the same trust model banks use.
• Built-in 14-type PII scanner strips emails, social security numbers, and similar data before anything leaves your machine — federation stays private.
• Native tools expanded from 87 to 314 in this release — CI checks, security scans, cloud deploys, observability integrations all added.
• Adaptive backpressure: when one agent gets too busy, others automatically slow down to prevent timeouts on long jobs.
• Installs into Claude Code with a single line; agents continue working across machine restarts.

How Federation Works

Federation is the flagship feature of v3.6.12. Before this release, RooFlow was confined to a single machine — your local Claude Code install orchestrated a team of agents, but that team couldn't collaborate with agents running on another computer. Federation changes this by letting two or more RooFlow instances establish a trusted connection using mTLS mutual authentication and ED25519 cryptographic keys. Once connected, agents can share tasks and pass outputs between machines. The built-in PII scanner strips sensitive data (emails, social numbers, financial identifiers — 14 types total) before anything crosses the wire. The practical example given: two financial institutions sharing fraud detection signals without sharing customer data.

What AgentDB Means in Practice

Previous multi-agent setups shared context through conversation threads or files — fragile, slow, and session-scoped. RooFlow's AgentDB is a specialized vector store built for agent-to-agent knowledge sharing. Agents write findings, decisions, and context to a shared memory that every other agent in the swarm can query in milliseconds. The 150–12,500x speed claim over basic search comes from the vector indexing approach. More practically: an agent that finished a task last Tuesday can have its findings available to a fresh agent starting a related task today, without any manual handoff or file passing.

Related on OpenClawDatabase

• Claude Cowork Setup Guide — getting started with Claude's native agent platform
• OpenClaw Skills Guide — custom automation layers for OpenClaw
• Hermes MCP Tools — tool integrations for the Hermes agent platform
• Compare Agents — side-by-side breakdown of all major platforms