Ruflo vs Claude Code subagents
Claude Code subagents are built in: the main agent hands a focused task to a helper with its own context and gets a summary back. Ruflo is a third-party layer installed around Claude Code or Codex that adds swarm topologies, on-disk vector memory, hooks and an MCP server. Start with subagents; add Ruflo for what they lack.
Facts side by side
| Fact | Ruflo | Claude Code subagents |
|---|---|---|
| What it is (source) | A third-party MIT-licensed orchestration layer (formerly Claude Flow) installed around Claude Code or Codex CLI: hooks, an MCP server, swarms and vector memory. | A built-in Claude Code feature: specialized assistants that run in their own context window inside one session and return a summary. |
| Setup (source) | npx ruflo@latest init writes CLAUDE.md, hooks and settings and registers the Ruflo MCP server; individual Claude Code plugins are a lighter path. | None for built-in subagents; custom ones are Markdown files in .claude/agents/ or ~/.claude/agents/. |
| Coordination model (source) | Swarms with a chosen topology (hierarchical, mesh, ring, star, hybrid, adaptive) and a hive-mind mode with queen agents and consensus algorithms. | The main agent delegates by matching the task to a subagent's description; results return to the caller. |
| Memory across sessions (source) | AgentDB with an HNSW index, a SQLite store and local embeddings on disk, synced with Claude Code's auto-memory files. | A subagent starts with its own context; persistence is whatever the session and project files provide. |
| Limits (source) | The guide recommends hierarchical swarms of at most eight agents for coding tasks. | By default 20 concurrent subagents per session and nesting up to three layers below the main conversation, both configurable. |
| Hosts (source) | Claude Code and OpenAI Codex CLI (init --codex, init --dual). | Claude Code only. |
| Footprint (source) | Default install is about 340MB with ML and embedding dependencies; about 45MB with --omit=optional. | No extra install. |
| Maintainer (source) | ruvnet (independent project, not affiliated with Anthropic). | Anthropic, documented in the official Claude Code docs. |
Choose Ruflo if
- You want shared memory that persists across sessions in a local vector store, beyond what project files give you.
- You run both Claude Code and Codex CLI on one project and want one layer to share memory and skills between them.
- You accept a large add-on: hooks, a background daemon, several hundred MCP tools and a release every few days.
Choose Claude Code subagents if
- You need focused helpers for search, review or verification whose output should stay out of the main context.
- You want behavior documented and supported by the harness vendor, with no extra install.
- You want to reuse the same role definitions later as agent-team teammates, which the official docs support.
Migration notes
The two can coexist: Ruflo installs into a Claude Code project and the built-in subagents keep working. To try Ruflo with the least change, add individual plugins through the Claude Code plugin marketplace (/plugin marketplace add ruvnet/ruflo); these add slash commands and agent definitions without hooks. The full CLI install rewrites CLAUDE.md, adds hooks and settings and registers an MCP server, so commit or back up the workspace first. Many package names, commands and the MCP server key still use the old claude-flow name. Check the open issues cited on the Ruflo tool page before relying on it. To go back, remove the hooks, the MCP registration and the .claude-flow/ folder; custom subagent files in .claude/agents/ are unaffected.
FAQ
Does Ruflo replace Claude Code subagents?
No. Ruflo sits around Claude Code: hooks and its MCP server route tasks and spawn helper agents, while Claude Code keeps running as usual with its own subagents.
Is Ruflo an Anthropic product?
No. It is an independent MIT-licensed project by ruvnet, formerly named Claude Flow.
Can Ruflo use Claude Code agent teams?
Its docs say it can switch on Claude Code's experimental Agent Teams feature and attach hooks to teammate-idle and task-completed events. Agent teams themselves are documented by Anthropic as experimental and off by default.
Which costs fewer tokens?
The official docs describe subagents as the lower-cost option because results are summarized back to the main context. No official token figures exist for Ruflo swarms; more agents mean more model calls.