Ruflo wraps Claude Code and Codex in plugins, MCP tools and hooks, and the Claude Code side gets the most

Four Claude Code plugins, and the mods run with your account permissions

The primary install path is the Claude Code plugin marketplace, and it requires Claude Code 2.1.287 or later. Inside Claude Code you add the marketplace, install three plugins, reload, and open Ruflo:

/plugin marketplace add ruvnet/ruflo
/plugin install ruflo-core@ruflo
/plugin install ruflo-console@ruflo
/plugin install ruflo-mods@ruflo
/reload-plugins
/ruflo

ruflo-core provides the foundation tools, ruflo-console opens the agent cockpit, and ruflo-mods adds routing and policy enforcement inside Claude Code. That last plugin carries the main risk: mods run with your account's permissions, and the project itself tells you to review their code before installing. On a work account with access to production repositories or secrets, treat ruflo-mods like any other code you would run under your own credentials, and read it first.

Codex and other MCP clients get the tools, not the console mods

For Codex or another MCP-enabled tool, Ruflo sets up the project through npx and then serves its tools over a local stdio MCP server. The setup wizard runs in the project terminal, and the server command is what you point your client at, with npx as the command and -y, ruflo@latest, mcp, start as the arguments.

npx ruflo@latest init wizard
npx -y ruflo@latest mcp start

The gap is stated plainly: MCP exposes Ruflo tools to the client, while the console and mods are Claude Code integrations, and hook support depends on the client. Since the automatic task routing after init runs through hooks, a Codex user should not expect the same background routing and learning that a Claude Code user gets. Check what hooks your client supports before planning around that behaviour.

init hands routing to hooks so you never touch 314 MCP tools

Ruflo exposes 314 MCP tools and 26 CLI commands, which is a lot of surface to learn. The intended answer is one npx ruflo init, after which you keep using Claude Code normally and the hooks system routes tasks, learns from successful patterns and coordinates agents in the background.

Two caveats sit next to that promise. The learning loop diagram is labelled conceptual: local vector retrieval and contrastive updates can run without an LLM, but embeddings and configured learning modules are still required. The architecture flow, from user through CLI or MCP, router, swarm, agents and memory to LLM providers, is also marked as a conceptual flow, not a live execution trace. The practical consequence is that routing decisions happen out of sight, so when a task lands on an unexpected agent you will need the console logs to see why.

The terminal console runs on its own, but the demo skips real execution

The cockpit does not require Claude Code. npx ruflo tui opens it in a plain terminal, and adding --plain prints one frame, which suits scripts. It shows Overview, Cost, Swarm, Events and Missions pages, plus an optional AI page, and tracks workflows, agents, models, tokens and cost.

The promotional animation is a 38 second, 3.84 MB SVG holding 456 actual renderer frames of the Orbital view and Theme Studio, including boot checks, six palettes, saved themes, JSON import and resizing. Its own caption says native host and AI provider execution are not demonstrated. So the demo shows the interface, not an agent swarm doing work, and you cannot judge orchestration quality from it.

The package is still published as claude-flow with mismatched agent counts

Ruflo began as Claude Flow; the first commit is dated June 2, 2025, and a June 10 snapshot added orchestration, task scheduling, memory and terminal modules. The rename has not reached every file. The root package.json is named claude-flow, version 3.71.0, exposes a bin called claude-flow, and its homepage and issue tracker point to ruvnet/claude-flow.

The numbers also disagree. The README pitches 100+ specialized agents, while the package.json description says 60+. Neither figure is reconciled anywhere visible, so quote neither as a spec. When searching for older issues or documentation, look under both names, since links still resolve to the claude-flow repository.

Grok Build writes only inside the project and can dry-run first

Grok support has its own init flag. npx ruflo init --grok writes the project host surface, meaning the .grok/ directory, the team bus and the operator guide, and nothing outside the project. Adding --dry-run previews what would be written. The Grok operator guide under v3/@claude-flow/cli/templates/grok/docs/ covers setup, the spawn_subagent contract, an optional status row, and what to re-check when Grok updates. That last item is the warning: the integration depends on Grok behaviour that can change, so plan to re-verify after each Grok update rather than assuming a working setup stays working.

Cargo.toml excludes gastown-bridge to avoid a workspace root clash

Ruflo is primarily TypeScript, but a root Cargo.toml lists two Rust members: v3/crates/ruflo-federation-peer and v3/crates/ruflo-agntcy. The v3/plugins/gastown-bridge directory is deliberately excluded because it declares its own nested WASM-focused workspace for wasm/gastown-formula-wasm and wasm/ruvector-gnn-wasm. Cargo refuses to resolve any member of a parent workspace while a listed member is itself a workspace root, failing with "multiple workspace roots found in the same workspace". The exclusion keeps cargo check and cargo test working on the other crates. If you need the gastown bridge, build it from its own directory, not from the repository root.

Three minor releases landed on one day, so @latest moves fast

On 2026-11-01 alone the project shipped v3.69.0, which lets you edit what agents remember through agentdb 3.3 CRUD, v3.70.0, which added governed context delegation with checked citations as an opt-in, and v3.71.0, which brought automatic context delegation in ruflo-mods 0.15.0, also opt-in. The last push was the same day. Every documented npx command uses ruflo@latest, which means two runs a few hours apart can pull different behaviour. For a shared team setup, pin a specific version in your MCP configuration and upgrade on purpose.

Editorial conclusion

Ruflo fits Claude Code users who want swarms, memory and routing hooks and are prepared to read the mods code before installing it, since mods run with their account permissions. Codex and other MCP users get the tools but not the console mods or guaranteed hooks. Before adopting it, confirm your Claude Code is 2.1.287 or later and pin a release instead of running @latest.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes