🛡️ Swarm

Engine-agnostic multi-agent orchestration. 245 agents. One command.

curl -fsSL https://raw.githubusercontent.com/Anasabubakar/agent-swarm/main/install.sh | bash

Then in any project folder:

swarm "Build a landing page"

That's it.

What Happens

swarm "Build a todo app with React"
  ↓
📁 Creates isolated workspace: projects/build-a-todo-app-with-react-20260401/
❓ Questionnaire asks clarifying questions
📐 Planner designs the implementation
🚀 Frontend + Backend + QA agents build in parallel
🐛 Debugger fixes any failures
🔍 Tech Lead reviews everything
  ↓
✅ Done. Check your workspace.

Install

Option 1: Curl (Recommended)

curl -fsSL https://raw.githubusercontent.com/Anasabubakar/agent-swarm/main/install.sh | bash

Option 2: NPM

npm install -g @anasabubakar/swarm

Option 3: Manual

git clone https://github.com/Anasabubakar/agent-swarm.git ~/.swarm
ln -sf ~/.swarm/bin/swarm /usr/local/bin/swarm

Prerequisites

• Python 3.8+
• Git
• At least one AI engine: kilo, claude, gemini, codex, aider, or cursor-agent

Usage

# Interactive — just shows help
swarm

# Give a goal — orchestrator handles everything
swarm "Build a landing page for TeenovateX Labs"

# Use a specific engine
swarm -e gemini "Create a REST API"

# Run a single agent
swarm -a frontend-dev "Create a responsive navbar"

# List what's available
swarm --list-engines
swarm --list-agents

What's Inside

Category	Count	What
🤖 Agents	245	Frontend, backend, devops, security, design, QA, PM, and 200+ more
🛠️ Skills	239	TDD, code review, debugging, deployment, security scanning
📋 Commands	125	Slash commands for common workflows
📏 Rules	77	Language-specific coding standards (TS, Python, Go, Rust, etc.)
🔧 Engines	11+	Claude, Gemini, Kilo, Codex, Cursor, Aider, Windsurf, Copilot, etc.

How It Works

5-Phase Workflow

1. Questionnaire — Asks clarifying questions before building
2. Planner — Designs architecture and task dependencies
3. Execute — Dispatches specialist agents in parallel
4. Debug — Automatically fixes failures
5. Ship — Tech Lead final review

Workspace Isolation

Each goal creates its own project folder. The swarm folder stays untouched.

projects/
└── your-goal-timestamp/
    ├── src/                    ← agents build here
    ├── tests/
    ├── .swarm-meta.json        ← tracking metadata
    └── .swarm-report.json      ← final report

Safety

Commands are classified:

• ✅ Safe — Auto-approve (ls, cat, git status)
• 🟡 Moderate — Auto-approve, logged (npm install, git commit)
• 🔴 Dangerous — Needs approval (rm, deploy, git push --force)
• 🚫 Blocked — Never allowed (rm -rf /, shutdown)

Self-Healing

If an agent fails:

1. Retry — Same agent, exponential backoff
2. Reassign — Swap to backup agent
3. Simplify — Break task into smaller pieces
4. Fallback — Switch to different engine
5. Escalate — Flag for human review

Engine Support

Works with ANY CLI agent that accepts a prompt:

Engine	Status
Claude Code	✅
Gemini CLI	✅
Kilo Code	✅
Codex	✅
Cursor Agent	✅
Aider	✅
Windsurf	✅
GitHub Copilot	✅
OpenCode	✅
Qwen Code	✅
Any custom CLI	✅ (via --engine generic)

Adding Your Own Agent

Drop a .md file in agents/:

# My Agent

## Role
You are a [specialty]. You ONLY do [thing].

## Constraints
- DO [things]
- DON'T [things]

Source Repos (Integrated)

• agency-agents — 185 specialist agents
• everything-claude-code — 147 skills, 36 agents, 68 commands
• superpowers — TDD, parallel dispatch, code review
• get-shit-done — 18 agents, meta-prompting
• claude-mem — Persistent memory system

License

MIT — Copyright (c) 2026 Anas Abubakar