Ruflo (formerly Claude Flow): multi-agent teams for Claude Code

• Sets up Ruflo in your project: settings, agents, skills, memory and hooks

  Run it in a project folder. It writes .claude/, .claude-flow/, CLAUDE.md and MCP config. Read the safety section first.

  Try: “npx ruflo@latest init wizard”

  Write · npx ruflo init

• Install Ruflo as Claude Code plugins, starting with ruflo-core

  Adds the ruvnet/ruflo marketplace, then you pick plugins. ruflo-core is the foundation and carries its own MCP server.

  Try: “/plugin install ruflo-core@ruflo”

  Plugin marketplace

• A cockpit inside Claude Code for workflows, agents, tokens and cost

  Open it with /ruflo after installing the ruflo-console plugin. You can create tracked missions from it.

  Try: “Create a mission to add a dark mode toggle. Show me the mission and its status.”

  ruflo-console (/ruflo)

• Starts hierarchical or mesh teams of typed agents like coder, reviewer and tester

  The swarm and agent tools let the assistant spawn several agents that split one task.

  Try: “ruflo swarm init --topology hierarchical --max-agents 8”

  Runs code · Swarms and agents

• Stores and searches what earlier sessions learned

  Memory tools save and look up notes with a vector index, so a new session can reuse old findings.

  Try: “ruflo memory search --query "auth patterns"”

  Write · Memory

• Runs scripts on prompts, commands and edits to route work and learn patterns

  Hooks watch what happens in a session and pick which agent or model should handle a task.

  Runs code · Hooks and routing

• Checks Node, npm, MCP servers, memory and API keys, then repairs problems

  Run it when something does not load.

  Try: “npx ruflo doctor --fix”

  Write · ruflo doctor --fix

• Lists, installs and toggles Ruflo plugins from the terminal

  The plugins command group handles list, install, uninstall, enable and disable. See ruflo-core, ruflo-swarm and ruflo-security-audit for examples.

  Plugins list

• Queen-led agent coordination; has an option to launch Claude without permission prompts

  It spawns a queen agent and workers on one objective. Its code declares --dangerously-skip-permissions as on by default; whether that default takes effect is not confirmed, so pass --no-auto-permissions to keep prompts.

  Runs code · hive-mind

• Removes files Ruflo created; previews first

  Dry run by default. Add --force to delete.

  Delete · ruflo cleanup

• Runs Claude Code and headless Codex workers on one feature

  Part of the @claude-flow/codex package.

  Try: “npx @claude-flow/codex dual run --template feature --task "Add user auth"”

  Runs code · Codex dual mode

Watch out for

• Hooks run on every command and edit

  init and the ruflo-core plugin register hooks on prompts, Bash, file edits and sessions that run Node scripts. Read .claude/settings.json after init and try it in a throwaway repo first. github.com

• It pre-approves its own commands

  init allows npx claude-flow commands and every claude-flow MCP tool without asking. Remove those allow rules if you want a prompt each time. github.com

• It turns on experimental agent teams

  init sets CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Claude Code docs say this launches named subagents as teammates, which use more tokens. code.claude.com

• Can write outside your project

  A contributor reported that init adds a block to ~/.claude/CLAUDE.md and registers ruflo in ~/.codex/config.toml when Codex is installed. Check those files after running it. github.com

• Mods are not sandboxed

  The README says ruflo-mods runs with your account's permissions and asks you to read the code first. Skip mods unless you have. github.com

• hive-mind can skip permission prompts

  hive-mind can launch claude with --dangerously-skip-permissions. Its code declares that option on by default; whether the default takes effect is not confirmed. Pass --no-auto-permissions to keep prompts, and use it only in a disposable environment. The user guide also lists that flag as an optional setup step. github.com

• Past supply-chain and injection reports

  Users reported an obfuscated preinstall script in versions 3.1.0-alpha.55 to 3.5.2 and hidden instructions in tool descriptions. The maintainer says both were removed by 3.5.40, so use a current version. github.com

• Security scanners flag the skill

  skills.sh shows Socket: Fail and Snyk: Warn for this skill. skills.sh

• Daemon and co-author trailer are opt-in

  The background daemon does not start on its own, and the Co-Authored-By trailer is added only with --attribution. init also denies reading .env files. github.com

Security advisories

In their words