Direct answer
Claude Flow is now called Ruflo. It is a third-party, open-source agent meta-harness from Ruvnet for Claude Code and Codex. It is not an Anthropic product. Ruflo adds coordinated agents, shared memory, hooks, an MCP server, plugins, and background processes around a coding agent. Test it in a disposable repository before giving it production code, credentials, or broad command access.
Summary
- Name
- the project formerly known as Claude Flow or claude-flow now uses Ruflo as its current name and package.
- Publisher
- Ruvnet maintains the public MIT-licensed repository. The name does not make it part of Claude or Anthropic.
- Purpose
- it wraps Claude Code, Codex, and other model providers with routing, multi-agent coordination, persistent memory, hooks, plugins, and MCP tools.
- Install choices
- a Claude Code plugin adds a narrow command surface; the full CLI changes project files and registers a broader runtime.
- Current requirement
- both `ruflo` and legacy `claude-flow` packages declare Node.js 20 or newer.
- Decision
- use it only when measured coordination benefits outweigh the extra tools, state, permissions, cost, and review burden.
Understand the Ruflo name change
Search results, old tutorials, package names, and repository links use several names. The original `ruvnet/claude-flow` repository now redirects to `ruvnet/ruflo`. Its current README says Claude Flow is now Ruflo. The npm registry currently publishes both `ruflo` and `claude-flow` at version 3.36.0, updated 10 August 2026. Use the current repository as the source of truth rather than copying an older alpha command.
Ruflo can run around Claude Code, but Anthropic does not publish or support it. That distinction matters for security review, support, upgrades, compatibility, and claims about performance. Treat statements in the Ruflo README as project-maintainer claims unless an independent source reproduces them.
What Ruflo adds
A coding model can write and reason about code. Ruflo supplies an execution layer around it. The layer selects agent definitions, routes tasks, coordinates workers, and retrieves stored patterns. It also exposes MCP tools and runs event hooks.
- Agents
- role definitions for coding, testing, review, architecture, documentation, and other work.
- Swarms
- hierarchical, mesh, or adaptive coordination patterns for several workers.
- Memory
- indexed project state and retrieved patterns that can persist beyond one conversation.
- MCP
- a server that exposes a large tool surface to compatible coding agents.
- Hooks
- event-driven routing, checks, memory retrieval, and background behavior.
- Plugins
- smaller bundles of commands, skills, and agent definitions for a selected capability.
- CLI and daemon
- terminal commands and background processes for orchestration outside a normal chat.
More agents do not guarantee a better result. Parallel workers can repeat the same mistake, conflict, spend more tokens, or create a larger diff than one focused agent. Evaluate task success, defects, review time, cost, and recovery rather than agent count.
Choose the narrow or full setup
The project documents two install paths with very different effects. Read the current diff for the path you choose. Do not treat the plugin and full CLI as equivalent.
- Claude Code plugin
- installs selected slash commands, skills, and agent definitions. The project says this path does not register the Ruflo MCP server or add the full memory and swarm tool surface.
- Full Ruflo CLI
- the current initializer can write project instructions, configuration directories, helper files, and settings. It can also register MCP, hooks, and a daemon.
- Hosted interfaces and federation
- these add network, identity, data-flow, and service dependencies. Evaluate them separately from a local plugin trial.
Compare the current plugin and CLI paths in the Ruflo README before running either one.
Check prerequisites and trust boundaries
- Runtime
- install a supported Node.js 20 or newer release and confirm `node` and `npx` resolve to the expected binaries.
- Host agent
- install and authenticate the supported Claude Code or Codex surface you intend to use.
- Repository
- start with a disposable test project under version control and a clean working tree.
- Permissions
- decide which files, commands, network destinations, repositories, and credentials the harness may reach.
- Budget
- cap model, API, cloud, and background-agent use before enabling parallel work.
- Recovery
- keep a known-good commit, backup, process list, MCP configuration copy, and a way to stop background services.
Test Claude Flow safely
- Confirm that `github.com/ruvnet/ruflo` and the npm publisher metadata are the sources you intend to trust.
- Read the current README, status, security policy, package manifest, install script, open issues, and recent release changes.
- Record the exact version and integrity information you approved. Avoid an unreviewed floating `latest` version in repeatable team setup.
- Prefer the narrow plugin path for an initial capability test. If you need the full CLI, inspect its initializer before execution.
- Run the trial in a disposable repository without production secrets, customer data, cloud credentials, signing keys, or open administrator sessions.
- Capture the file tree, Git status, MCP configuration, processes, and network policy before installation. Compare them after setup.
- Start one bounded, reversible task. Limit agents, commands, runtime, model spend, allowed paths, and expected output.
- Review every diff and command log. Run independent tests, type checks, linting, security checks, and a clean build.
- Stop background processes and remove access when the trial ends. Keep the tool only if measured outcomes beat the simpler workflow.
The README offers a command that downloads a shell script and pipes it into a shell. A safer review process downloads the script, checks the exact bytes and source, and runs it only after approval. An interactive initializer can still change many files, so it also needs review.
Use a bounded orchestration workflow
- Define one outcome, acceptance criteria, excluded actions, time limit, token budget, and stopping condition.
- Choose the fewest roles needed. A coder and reviewer may be enough; a large swarm can add coordination cost without new evidence.
- Give each worker a distinct responsibility, file boundary, source of truth, and required handoff artifact.
- Keep memory namespaced to the test project. Do not mix customer, personal, or unrelated repository context.
- Require approval before package installation, external messages, publication, money movement, deletion, deployment, merge, or credential use.
- Verify the final repository state with tools outside the orchestrated agent session.
- Compare success rate, defects, elapsed time, model spend, changed lines, review time, and rollback effort with one-agent work.
Treat project status as evidence, not a guarantee
The project publishes test counts, capability inventory, recent fixes, deferred work, and verification coverage. It also says some per-capability witness and functional verification remains in progress. That disclosure is useful. A passing project suite does not prove your installation, configuration, plugins, prompts, dependencies, or generated code are safe.
The current public documents also need reconciliation. Registry metadata showed version 3.36.0 on 11 August 2026, while the status snapshot described an older release and the security policy listed only 3.5.x as supported. Resolve version and support questions with the maintainer before production adoption.
Review Ruflo's current status, test baseline, shipped fixes, deferred work, and verification notes.
Review the expanded attack surface
A meta-harness can reach more resources than a read-only chat. It may edit files, run commands, store memory, call models, connect tools, or operate a browser. Other configurations can use GitHub, start workers, or communicate across machines. Each capability needs an owner and a policy.
- Supply chain
- review package ownership, dependency changes, release provenance, integrity, install scripts, and plugin publishers.
- Secrets
- keep credentials outside indexed memory and allowed paths. Use short-lived, least-privilege test credentials when access is unavoidable.
- Prompt injection
- treat repository files, issues, webpages, tool output, and retrieved memory as untrusted instructions.
- Commands
- use allowlists, sandboxing, timeouts, resource limits, and human approval for impactful actions.
- Memory
- document what is stored, where it lives, how it is encrypted, who can retrieve it, and how deletion is verified.
- Federation
- confirm identity, encryption, trust changes, outbound filtering, logs, revocation, and incident response before cross-machine use.
- Generated code
- inspect the diff and verify behavior, security, accessibility, performance, dependencies, and maintainability.
Check Ruflo's current supported-version and vulnerability-reporting policy.
Ruflo or native Claude Code?
- Use native Claude Code for a focused task, built-in subagents, project instructions, hooks, skills, MCP, or direct review when those features meet the need.
- Evaluate Ruflo when repeated work needs shared memory, explicit routing, several specialized workers, cross-session coordination, or a reusable orchestration policy.
- Use a normal script or CI job when the workflow is deterministic and does not need model judgment.
- Avoid extra orchestration when the team cannot review the generated work, govern credentials, measure cost, or recover from broad changes.
Compare the narrower built-in worker model in our Claude Code subagents guide.
Costs beyond the open-source license
The Ruflo source is MIT licensed, but the full workflow is not cost-free. Claude Code, Codex, third-party models, cloud agents, hosted infrastructure, storage, observability, and network services can create charges. Parallel workers can multiply model use. Budget the human time for setup, review, failures, upgrades, and incident response as well.
Plan removal before adoption
The full initializer can touch project instructions, settings, hook configuration, MCP registration, memory directories, helpers, and background services. A package uninstall alone may not restore the previous state. Use the captured baseline and version control to identify every change. Stop processes, remove registrations, revoke credentials, delete test memory according to policy, and rerun the original checks.
Frequently asked questions about Claude Flow
No. Ruflo is a third-party open-source project maintained under the Ruvnet GitHub organization. It can integrate with Claude Code, but Anthropic does not publish it.
Yes in the current project history. The repository redirects from the old Claude Flow name, and the README says the project is now Ruflo. Old commands and alpha tutorials may describe a different surface.
The public source uses the MIT license. Model subscriptions, API tokens, cloud resources, storage, and third-party services can still cost money. Several agents may consume more than one focused session.
Start with the host coding agent and version control first. Add a narrow Ruflo plugin only when you can explain the missing capability, inspect the changes, cap access, and recover the repository.
That depends on your version, configuration, plugins, threat model, support needs, and evidence. Reconcile the current support documents, run a controlled evaluation, and require security and platform owners to approve production access.
Sources and verification
- Ruvnet Ruflo repository and README
- current name, architecture, install paths, changed files, plugins, commands, stated capabilities, and license.
- Ruflo status page
- versioned test baseline, capability inventory, shipped hardening, deferred work, and verification limitations.
- Ruflo user guide
- detailed commands and configuration reference.
- Ruflo security policy
- supported-version statement and vulnerability-reporting process.
- npm registry metadata checked 11 August 2026
- current `ruflo` and `claude-flow` versions, Node requirement, repository, and license.
- Independent current guides from Analytics Vidhya, Tosea, and Nahornyi AI Lab
- search coverage and reader questions, not proof of maintainer claims.