Ruflo drops an execution layer into Claude Code, turning agents into swarms with memory and routing
Synopsis
Ruflo enters Claude Code as a plugin set that adds the execution layer around agents: tools, memory, loops and policy control. Previously agents in Claude Code ran without a shared coordination layer for task routing or cross-session memory. Ruflo connects through three plugins, core tools, a console and runtime mods, and can also run as an MCP server for other MCP-enabled clients. The console and mods are Claude Code integrations, and hook support depends on the client.
Interpretation
Claude Code users can now install Ruflo to move agents from running alone to collaborating as swarms that keep memory across sessions. Agents in Claude Code previously lacked a unified coordination layer, so task routing, cross-session memory and policy enforcement had to be assembled separately. The install path requires Claude Code 2.1.287 or later and adds the core, console and mods plugins from the marketplace before reloading.
Ruflo separates the model from the execution layer: the model writes, while the layer supplies tools, memory, loops, sandboxes and controls so agents can actually do work. That split maps to a flow running from the CLI or MCP entry through a router into swarms, agents and memory, then out to model providers. The flow is described as a conceptual flow rather than a live execution trace, and memory feeds useful experience back into routing.
Teams not using Claude Code can take the MCP path, starting a local stdio server with npx to expose Ruflo tools to MCP-enabled clients. The console and runtime mods are Claude Code integrations, and hook support varies by client. Configuration uses the command npx with arguments -y ruflo@latest mcp start; the console also runs standalone via npx ruflo tui, with --plain printing one frame for scripts.
Perspective
Intended for developers on Claude Code 2.1.287 or later who want multi-agent coordination and memory in their coding workflow. Installation goes through the plugin marketplace, with core, console and mods supplying foundation tools, an agent cockpit, and routing and policy enforcement. Teams outside Claude Code can start an MCP server with npx, and the console also runs standalone.
Runtime mods run with the account's permissions, and the source advises reviewing their code before installing. The console and mods are Claude Code integrations, and hook support depends on the client. The architecture flow is explicitly conceptual rather than a live execution trace; local vector retrieval and contrastive updates can run without an LLM but still require embeddings and configured learning modules.