Claude-Flow — Multi-Agent Orchestration for Claude Code

Introduction

Claude-Flow is an orchestration layer that turns a single Claude Code instance into a coordinated swarm of specialized agents. It ships 64 agent archetypes (architect, reviewer, tester, refactorer, security, and more), a shared SQLite memory layer, and modes for parallel and hive-mind execution — letting one task fan out across many roles without you wiring up the messaging yourself.

What Claude-Flow Does

• Coordinates multiple Claude Code agents on one task, each with a focused role.
• Provides 64 pre-built agent profiles covering the full SDLC (planning, coding, review, testing, ops).
• Runs in two modes: swarm (parallel agents per subtask) and hive-mind (shared memory across agents).
• Uses a SQLite-backed memory store so agents see each other findings and avoid duplicate work.
• Generates an orchestration plan from a single high-level prompt — no manual fan-out script required.

Architecture Overview

Claude-Flow is a Node.js CLI distributed via npm. On init it scaffolds a project with agent registries, memory tables, and config. Each agent is a Claude Code subprocess plus a role-specific system prompt; the orchestrator spawns them concurrently, watches stdout, persists outputs to SQLite, and resolves dependencies between agents based on the plan it generated up front.

Self-Hosting & Configuration

• npx claude-flow@alpha init writes a .claude-flow/ directory with agent registry, memory schema, and run config.
• Agents are configurable Markdown — add or edit profiles to fit your stack.
• SQLite memory file is local; commit it (or ignore it) like any other build artifact.
• Pin a version: npx claude-flow@<tag> — the project is on an alpha release cadence.
• Telemetry is opt-in; without env vars set, runs stay local.

Key Features

• Out-of-the-box swarm orchestration — no LangGraph or AutoGen wiring required.
• 64 specialized agents you can mix and match for any task profile.
• Hive-mind mode shares memory so parallel agents stay coherent.
• Built specifically for Claude Code; uses its tool ecosystem rather than reimplementing it.
• Plan-first execution: the orchestrator decides who runs what and in what order before any agent fires.

Comparison with Similar Tools

• Single Claude Code session — simplest, but no parallel role specialization.
• AutoGen / CrewAI — model-agnostic orchestration frameworks; heavier, generic, more wiring.
• LangGraph — graph-based agent runtime; lower-level than Claude-Flow, more flexible.
• OpenAI Swarm — similar concept on the OpenAI side; Claude-Flow is the Claude-native sibling.
• Manual subagent calls in Claude Code — fine for ad hoc; Claude-Flow gives you a registry, memory, and plans.

FAQ

Q: Do I need API credits separate from Claude Code? A: No — agents are Claude Code subprocesses, so they use whatever auth your CLI has.

Q: Is hive-mind required? A: No. swarm is the default; hive-mind is a stronger coordination mode for tasks that need shared state.

Q: Can I use only a few of the 64 agents? A: Yes. Edit the agent registry in .claude-flow/ and remove the ones you do not want spawned.

Q: What happens if an agent errors out? A: The orchestrator marks it failed and routes around it; downstream agents that depended on its output get a structured failure they can retry.

Sources

• https://github.com/ruvnet/claude-flow
• https://www.npmjs.com/package/claude-flow