Skip to content

zircote/claude-team-orchestration

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

48 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
swarm - Multi-agent orchestration for Claude Code

MIT License Claude Code Plugin Agent Teams Version RLM Pattern

Multi-agent orchestration plugin for Claude Code. Coordinate teams of agents with shared tasks, inter-agent messaging, and proven patterns for parallel reviews, pipelines, self-organizing swarms, and large-file analysis via the RLM pattern.

Official docs: Agent Teams

How Swarm Orchestration Works


Prerequisites

Before installing, make sure you have:

Claude Code CLI

Install and authenticate Claude Code. See the quickstart guide.

claude --version   # Verify installed, must be 1.0.33+

Enable Agent Teams

Agent teams are experimental and disabled by default. Enable them by adding this to your settings.json:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Or set the environment variable directly:

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

tmux (recommended)

Split-pane mode lets you see all teammates working side-by-side. It requires tmux.

macOS:

brew install tmux

Linux (Debian/Ubuntu):

sudo apt install tmux

Why tmux? Background agents need terminal multiplexing to run in visible, persistent panes. Without tmux, agents run in-process (invisible but functional). See docs/getting-started.md for details on display modes.

iTerm2 (optional, macOS only)

If you use iTerm2, install the it2 CLI for native split panes:

uv tool install it2
# Then: iTerm2 -> Settings -> General -> Magic -> Enable Python API

Installation

From GitHub

claude /plugin install https://github.com/zircote/claude-team-orchestration

From local directory

Clone the repo and load it directly:

git clone https://github.com/zircote/claude-team-orchestration.git
claude --plugin-dir ./claude-team-orchestration

Verify installation

Start Claude Code and run:

/help

You should see swarm skills listed under the swarm: namespace:

swarm:orchestrating
swarm:team-management
swarm:task-system
swarm:messaging
swarm:agent-types
swarm:orchestration-patterns
swarm:spawn-backends
swarm:error-handling
swarm:rlm-pattern
swarm:jsonl-log-analyzer

Troubleshooting setup

Problem Fix
Skills don't appear Restart Claude Code after installing
swarm: prefix missing Verify .claude-plugin/plugin.json exists in the plugin directory
Agent teams not working Check CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS is set to 1
Split panes not showing Run which tmux to verify tmux is installed

Quick Start

Here's a minimal example that spawns a parallel code review team with three specialists:

Create an agent team called "review" to review the src/ directory.
Spawn three reviewers in parallel:
- "security" using sdlc:security-reviewer — check for vulnerabilities
- "quality" using feature-dev:code-reviewer — check for bugs and performance
- "simplicity" using code-simplifier:code-simplifier — check for unnecessary complexity

Have each reviewer send findings to team-lead when done.
Synthesize all findings into a summary, then shut down the team.

What success looks like:

  1. Claude creates the team and spawns three agents
  2. Each agent reviews the code through its own lens (in parallel)
  3. Findings arrive as messages to the team lead
  4. The lead synthesizes a unified report
  5. The team shuts down cleanly

For more examples, see docs/patterns.md and skills/orchestration-patterns/examples/complete-workflows.md.


Recommended CLAUDE.md Configuration

To get the most out of swarm orchestration, add the following to your project's CLAUDE.md (or your personal ~/.claude/CLAUDE.md). This tells Claude to prefer parallel agent teams over sequential work whenever appropriate:

## Always use swarm orchestration patterns (TeamCreate, Task with team_name, SendMessage, TaskCreate/TaskUpdate) when work is best executed by parallel specialist agents.
This means:
  - Spawning teams with specialized teammates for parallelizable work
  - Using the task list for coordination and progress tracking
  - Leveraging SendMessage for inter-agent communication
  - Preferring concurrent execution over sequential when tasks are independent

Why this matters: Without this instruction, Claude may default to running tasks sequentially or using simple subagents instead of coordinated teams — even when the plugin is installed. Adding this to CLAUDE.md ensures Claude proactively reaches for team orchestration when it would be beneficial.

Where to put it:

  • Project-level (CLAUDE.md at your project root) — applies to all sessions in that project
  • Personal global (~/.claude/CLAUDE.md) — applies to all your projects

Skills Reference

All skills are available under the swarm: namespace.

Skill Command Description
Orchestrating swarm:orchestrating Primitives overview, lifecycle, quick reference. Start here.
Team Management swarm:team-management Create teams, spawn teammates, delegate mode, permissions, shutdown, cleanup
Task System swarm:task-system Create tasks, set dependencies, claim work, track progress
Messaging swarm:messaging Direct messages, broadcasts, shutdown requests, plan approvals
Agent Types swarm:agent-types Built-in and plugin agents, selection guide, capabilities
Orchestration Patterns swarm:orchestration-patterns 7 patterns: parallel specialists, pipelines, swarms, research+implement, plan approval, refactoring, RLM
Spawn Backends swarm:spawn-backends In-process, tmux, iTerm2 backends, auto-detection, configuration
Error Handling swarm:error-handling Common errors, quality gate hooks, known limitations, recovery strategies
RLM Pattern swarm:rlm-pattern Content-aware chunked analysis of large files and directories using RLM pattern
JSONL Log Analyzer swarm:jsonl-log-analyzer Schema-aware analysis of large JSONL log files with automated field discovery and jq recipes

Documentation

Guide Description
Getting Started End-to-end walkthrough from install to first team run
Agent Types Which agents are available and when to use each
Patterns Orchestration patterns with guidance on when to pick which
RLM Examples Copy-paste example prompts for every RLM mode
Concepts How teams, patterns, and RLM content processing work internally
Reference Primitives, message types, task fields, configuration, and API lookup tables
Troubleshooting Common failures and fixes

Attribution

This plugin is based on the Claude Code Swarm Orchestration Skill by @kieranklaassen. The original monolithic gist has been decomposed into modular, maintainable skills that can be independently loaded, updated, and extended.