Ruflo Agent Harness
A coordination, memory, and execution layer for multi-agent work in Claude Code and Codex.
Per-dimension scores and reasoning
Evidence shows security measures (e.g., AIDefence, input validation, path traversal prevention) and PII stripping in federation, but lacks user confirmation mechanisms and detailed least-privilege descriptions. Deductions: no explicit user confirmation process, insufficient permission management details.
Tests (e.g., arena tests) and CI workflows exist, but dependencies are numerous and some are alpha versions; failure messages are not detailed. Deductions: dependency availability unverified, failure messages insufficient.
Targets multiple scenarios (coding, testing, security, etc.) with a plugin system, but trigger precision and boundary descriptions are unclear. Deductions: trigger mechanism details insufficient, environment fit not fully documented.
Has README, install notes, license, but naming stability (e.g., claude-flow vs ruflo) and versioning/changelog are unclear. Deductions: naming inconsistency, missing changelog.
Output usability is high (CLI, MCP, Web UI), marginal value is evident, but cost-benefit is not quantified. Deductions: cost-benefit analysis missing.
Tests and audit docs exist, but claim-to-evidence mapping is incomplete, cross-source corroboration insufficient. Deductions: limited claim traceability, unclear fact-inference separation.
- Numerous dependencies, some alpha versions, need supply chain security attention.
- Naming inconsistency (claude-flow vs ruflo) may cause confusion.
- Lack of explicit user confirmation mechanisms, automated operations need caution.
What does this agent do, and when should you use it?
Ruflo is an agent meta-harness for Claude Code and Codex, delivered through a CLI, MCP server, and Claude Code plugins. Its full CLI installation creates workspace configuration and installs the MCP server, hooks, daemon, commands, skills, and specialized agents. The system routes work into swarms using hierarchical, mesh, or adaptive topologies, with AgentDB, HNSW, SONA, and ReasoningBank described as its memory and learning layer. It also includes a plugin marketplace, cross-machine federation, a hosted or self-hostable Web UI, and a goal-planning UI. It fits teams that want persistent context and coordinated execution around coding agents, rather than a lightweight slash-command-only plugin.
After npx ruflo@latest init wizard or init, Ruflo installs workspace integration and can be registered in Claude Code with claude mcp add ruflo -- npx ruflo@latest mcp start. Its router assigns work to specialized agents in a swarm; documented MCP operations include memory_store, swarm_init, and agent_spawn, while AgentDB with HNSW stores and retrieves memory. Its plugins cover operations such as reusable workflows, test generation, Playwright browser testing, documentation, risk analysis, security auditing, and observability. For remote collaboration, federation init, federation join, federation send, and federation status support federation workflows with documented identity, PII, trust, and audit processing.
- A Claude Code engineering team that wants coding, testing, review, and security work routed among specialized agents.
- A software team that needs project experience retained across sessions and retrieved through vector memory for later tasks.
- A developer integrating swarm coordination and persistent memory into Claude Code through an MCP server.
- A QA team using the browser plugin’s Playwright automation and the test-generation plugin to address missing tests.
- Teams collaborating across machines or organizations that need mTLS, ed25519, PII handling, and audit trails around agent-to-agent exchange.
- An internal platform team that wants a self-hosted chat UI with MCP calls or goal decomposition into executable agent plans.
What are this agent's strengths and limitations?
- The full CLI path combines an MCP server, hooks, daemon, commands, skills, and specialized agents rather than supplying plugin commands alone.
- It documents hierarchical, mesh, and adaptive swarm coordination connected to persistent AgentDB/HNSW memory.
- Its federation design explicitly includes mTLS, ed25519 challenge-response, PII handling, behavioral trust scoring, and audit trails.
- It offers Claude Code plugins, CLI/MCP delivery, a hosted Web UI, and a Docker self-hosting path, while naming Claude, GPT, Gemini, Cohere, and Ollama provider paths.
- The Claude Code plugin route is deliberately limited: it does not register MCP, so operations such as
memory_store,swarm_init, andagent_spawnare unavailable. - Full initialization writes
.claude/,.claude-flow/,CLAUDE.md, helpers, and settings into the workspace, which should be assessed against existing repository conventions. - Many capabilities are composed from plugins, MCP, background workers, and external model providers; the supplied material does not specify complete CLI runtime prerequisites or credential configuration.
- Federation introduces remote endpoints, identity verification, trust levels, and PII policy, increasing deployment and governance complexity over a local Claude Code workflow.
How do you install or deploy this agent?
For the full setup, run npx ruflo@latest init wizard; use npx ruflo@latest init for non-interactive setup, or npm install -g ruflo@latest for a global install. On macOS, Linux, WSL, or Git-Bash, the documented installer is curl -fsSL https://cdn.jsdelivr.net/gh/ruvnet/ruflo@main/scripts/install.sh | bash. For Claude Code slash commands only, run /plugin marketplace add ruvnet/ruflo, then install plugins such as /plugin install ruflo-core@ruflo and /plugin install ruflo-swarm@ruflo; this path does not register the MCP server or hooks. To register the full MCP service in Claude Code, run claude mcp add ruflo -- npx ruflo@latest mcp start.
How do you use this agent?
After a full initialization, use Claude Code normally; the documentation says its hooks route tasks, learn from successful patterns, and coordinate agents in the background. For direct integration, first register claude mcp add ruflo -- npx ruflo@latest mcp start, then use documented MCP capabilities such as memory_store, swarm_init, and agent_spawn. For cross-node work, start with npx claude-flow@latest federation init, join an endpoint with federation join <endpoint>, and send work with federation send. The hosted UI at flo.ruv.io is documented as usable without an account or API key; self-hosted UI source is under ruflo/src/ruvocal/ and includes a Dockerfile.
How does this agent compare with similar options?
The README contrasts “Claude Code Alone” with “+ Ruflo”: the former is described as having isolated collaboration, session-only memory, and manual coordination, while Ruflo adds swarm coordination, HNSW vector memory, SONA learning, routing, background workers, and multi-provider paths. This presents Ruflo as an enhancement layer around Claude Code, not a replacement for Claude Code itself.
FAQ
Does installing the Claude Code plugin provide the complete product?
Is Ruflo limited to Claude?
Do I need an API key to try the hosted UI?
Will it change my repository?
.claude/, .claude-flow/, CLAUDE.md, helpers, and settings. The Claude Code plugin path is marked as creating zero workspace files.