swarmgit 0.1.1

Multi-agent Git worktree orchestrator for AI coding agents

SwarmGit — Multi-Agent Git Worktree Orchestrator for AI Coding Agents

Run multiple AI coding agents in parallel using native git worktree. Each agent gets an isolated branch and working directory. Zero conflicts, maximum velocity.

SwarmGit is an open-source multi-agent orchestration CLI that solves the biggest pain point in AI-assisted software development: AI coding agents can't work in parallel on the same codebase without overwriting each other.

By leveraging native git worktree, SwarmGit gives every agent its own isolated workspace and branch. When agents finish, their changes are merged back into main automatically. It's like Kubernetes for AI coding agents — but for Git.

---

Table of Contents

• What is SwarmGit?
• The Problem (With a Real Example)
• Supported AI Coding Agents
• How It Works
• Installation
• Real-World Walkthrough
• CLI Reference with Examples
• Configuration
• Architecture
• Agent Adapters
• Contributing
• Roadmap
• License

---

What is SwarmGit?

SwarmGit is a Git worktree orchestrator designed specifically for the era of AI coding agents. It enables:

• Parallel agent execution — Run Claude Code, Aider, Codex, and OpenCode simultaneously on different parts of your project
• Automatic branch isolation — Each agent works in its own Git worktree linked to a dedicated branch
• Smart merge coordination — Completed agent branches are merged into main sequentially with conflict resolution
• Zero directory switching — Agents never run git checkout; each worktree is permanently bound to its branch
• LLM-powered task splitting — Automatically decompose large tasks into parallel subtasks using Claude or OpenAI

Use Cases

• Microservice scaffolding — Have one agent set up the API, another write auth, and a third write tests — all at the same time
• Refactoring at scale — Split a monolith refactor across multiple agents working in parallel
• Multi-language projects — Frontend agent builds the React app while the backend agent writes the FastAPI server
• Test generation — Spawn dedicated tester agents to write unit tests while the primary agent implements features

---

The Problem (With a Real Example)

Before SwarmGit — The Overwrite Nightmare

Imagine you want to build a REST API with authentication and tests. You have Claude Code and Aider installed. Here's what happens without SwarmGit:

# Terminal 1 — You ask Claude to build the API
$ claude "Build a FastAPI REST API with JWT auth"
# Claude starts editing app.py, models.py, auth.py...

# Terminal 2 — You ask Aider to write tests
$ aider "Write pytest unit tests for the auth module"
# Aider starts editing test_auth.py — but also touches app.py

# Result: Aider's changes to app.py overwrite Claude's changes.
# You lose hours of work and have to start over.

Workarounds people try:

• Copy directories manually — Lose git history, can't merge back easily
• Switch branches constantly — Confusing, error-prone, agents hate it
• Run agents one at a time — Defeats the purpose of having fast AI agents

After SwarmGit — Parallel Execution, Zero Conflicts

# One command. SwarmGit handles everything.
$ swarm run "Build a FastAPI REST API with JWT auth and tests"

# SwarmGit automatically:
# 1. Splits the task into subtasks
# 2. Creates isolated worktrees for each agent
# 3. Launches agents in parallel
# 4. Monitors progress
# 5. Merges results when done

Result: Claude works in .swarm/agent-1/, Aider works in .swarm/agent-2/. Both commit to their own branches. SwarmGit merges them into main when done. No overwrites. No lost work.

---

Supported AI Coding Agents

SwarmGit includes built-in adapters for the most popular AI coding agents. Adding a new agent requires only ~50 lines of Python.

Agent	CLI Command	Adapter Status	Git Integration
Claude Code	claude	✅ Built-in	Native Git commits
OpenCode	opencode	✅ Built-in	Native Git commits
OpenAI Codex	codex	✅ Built-in	Native Git commits
Aider	aider	✅ Built-in	Managed by SwarmGit (--no-git)
Custom Agent	User-defined	✅ Configurable	Via swarm.yml

---

How It Works

User runs:
  swarm run "Build a REST API with auth and tests"
        │
        ▼
  1. Orchestrator calls LLM (Claude/OpenAI) to split task into subtasks
        │
        ▼
  2. WorktreeManager creates one git worktree per subtask
     .swarm/agent-1/  →  branch: swarm/agent-1
     .swarm/agent-2/  →  branch: swarm/agent-2
     .swarm/agent-3/  →  branch: swarm/agent-3
        │
        ▼
  3. Each agent launches inside its worktree directory
     claude, opencode, codex, aider — whatever is configured
        │
        ▼
  4. AgentMonitor watches each agent process for completion
        │
        ▼
  5. MergeCoordinator merges DONE branches into main sequentially
        │
        ▼
  6. WorktreeManager cleans up worktrees

Directory Layout

/my-project/              ← main repo, always on branch: main
/my-project/.swarm/
    agent-1/              ← permanently on branch: swarm/agent-1
    agent-2/              ← permanently on branch: swarm/agent-2
    agent-3/              ← permanently on branch: swarm/agent-3

Critical principle: Agents never run git checkout. Each worktree is permanently mapped to its branch via git worktree. Your main repo stays on main at all times.

---

Installation

Requirements

• Python 3.10 or higher
• Git 2.30 or higher (for worktree support)
• At least one supported AI coding agent installed

Install from PyPI

pip install swarmgit

Install from Source

git clone https://github.com/rjben/swarm-git.git
cd swarm-git
pip install -e ".[dev]"

Verify Installation

$ swarm --version
swarm, version 0.1.0

$ swarm doctor
SwarmGit Doctor
────────────────────────────────────────
✅ Git 2.43.0 (worktree supported)
✅ Python 3.12.3
✅ claudecode found at /usr/local/bin/claude
✅ opencode found at /usr/local/bin/opencode
⚠️  codex not found (install: npm install -g @openai/codex)
⚠️  aider not found (install: pip install aider-chat)
────────────────────────────────────────
2 agents ready · 1 warning · 1 missing

---

Real-World Walkthrough

Let's build a real project together. We'll create a Python CLI tool with auth, tests, and documentation — all in parallel using 3 agents.

Step 1: Initialize the Project

$ mkdir url-shortener && cd url-shortener
$ swarm init .
Initialized empty Git repository in /home/user/url-shortener/.git/
Initialized SwarmGit project at /home/user/url-shortener
Run 'swarm task "your task"' to get started

$ ls -la
total 24
drwxr-xr-x 4 user user 4096 May  3 10:00 .
drwxr-xr-x 6 user user 4096 May  3 10:00 ..
drwxr-xr-x 8 user user 4096 May  3 10:00 .git
-rw-r--r--r-- 1 user user  655 May  3 10:00 swarm.yml
drwxr-xr-xr-x 2 user user 4096 May  3 10:00 .swarm
-rw-r--r--r-- 1 user user    8 May  3 10:00 .gitignore

Step 2: Configure Agents

Edit swarm.yml and set your API key:

splitter:
  provider: anthropic
  model: claude-sonnet-4-5
  api_key_env: ANTHROPIC_API_KEY

default_agent: claudecode
max_agents: 3

Step 3: Preview Task Splitting

$ swarm task "Build a URL shortener CLI in Python with Click, SQLite storage, unit tests, and a README"
Analyzing task...

Proposed subtasks:
  1. Create project structure, setup.py, and Click CLI skeleton with add/list/delete commands
  2. Implement SQLite database layer and URL shortening logic with hash generation
  3. Write pytest unit tests for database operations and CLI commands

3 agents will be spawned.
Accept? [Y/n]: y

Step 4: Run the Swarm

$ swarm run "Build a URL shortener CLI in Python with Click, SQLite storage, unit tests, and a README"
Task: Build a URL shortener CLI in Python with Click, SQLite storage, unit tests, and a README
Split into 3 subtasks

Spawned agent-1 (claudecode) -> swarm/agent-1
Spawned agent-2 (opencode) -> swarm/agent-2
Spawned agent-3 (codex) -> swarm/agent-3

Waiting for agents to complete...

Step 5: Monitor Progress (In Another Terminal)

$ swarm status --watch

SwarmGit — url-shortener
──────────────────────────────────────────────────────────
AGENT      TOOL          BRANCH              STATUS        LAST COMMIT
agent-1    claudecode    swarm/agent-1       ✅ done       "setup project structure and cli skeleton"
agent-2    opencode      swarm/agent-2       🟡 working    "implemented sqlite storage layer"
agent-3    codex         swarm/agent-3       ⏳ waiting    —
──────────────────────────────────────────────────────────
1 done · 1 working · 1 waiting · 0 failed

Step 6: Inspect Agent Work

# Check what agent-1 built without switching branches
$ swarm peek agent-1 --tree
.
├── setup.py
├── urlshortener/
│   ├── __init__.py
│   └── cli.py
└── README.md

# Read a specific file
$ swarm peek agent-1 --file urlshortener/cli.py
import click

@click.group()
def cli():
    """URL Shortener CLI"""
    pass

@cli.command()
@click.argument('url')
def add(url):
    """Add a URL and return a short code"""
    pass

# See the diff before merging
$ swarm diff agent-1
diff --git a/setup.py b/setup.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/setup.py
@@ -0,0 +1,15 @@
+from setuptools import setup, find_packages
+
+setup(
+    name='urlshortener',
+    version='0.1.0',
+    packages=find_packages(),
+    install_requires=['click'],
+    entry_points={
+        'console_scripts': [
+            'urlshortener=urlshortener.cli:cli',
+        ],
+    },
+)

Step 7: Merge Results

$ swarm merge
Merging agent-1 (swarm/agent-1) → main ...  ✅ clean
Merging agent-2 (swarm/agent-2) → main ...  ✅ clean
Merging agent-3 (swarm/agent-3) → main ...  ✅ clean

# Check the final result
$ git log --oneline -5
a1b2c3d Merge swarm/agent-3: Write pytest unit tests for database operations and CLI commands
e4f5g6h Merge swarm/agent-2: Implement SQLite database layer and URL shortening logic
i7j8k9l Merge swarm/agent-1: Create project structure, setup.py, and Click CLI skeleton
m0n1o2p Initial commit

$ ls -la
total 40
drwxr-xr-x 5 user user 4096 May  3 10:15 .
drwxr-xr-x 6 user user 4096 May  3 10:00 ..
drwxr-xr-x 8 user user 4096 May  3 10:15 .git
-rw-r--r--r-- 1 user user  655 May  3 10:00 swarm.yml
drwxr-xr-x 2 user user 4096 May  3 10:15 .swarm
drwxr-xr--r-- 1 user user    8 May  3 10:00 .gitignore
-rw-r--r--r-- 1 user user 1500 May  3 10:15 README.md
-rw-r--r--r-- 1 user user  300 May  3 10:15 setup.py
drwxr-xr-x 3 user user 4096 May  3 10:15 tests/
drwxr-xr-x 3 user user 4096 May  3 10:15 urlshortener/

Step 8: Clean Up

$ swarm clean
Removed worktree: /home/user/url-shortener/.swarm/agent-1
Deleted branch: swarm/agent-1
Removed worktree: /home/user/url-shortener/.swarm/agent-2
Deleted branch: swarm/agent-2
Removed worktree: /home/user/url-shortener/.swarm/agent-3
Deleted branch: swarm/agent-3
Cleaned up.

Done! You now have a complete URL shortener project built by 3 agents working in parallel, with zero conflicts.

---

CLI Reference with Examples

swarm init [project-name]

Initialize a new SwarmGit project.

$ swarm init my-api
Initialized empty Git repository in /home/user/my-api/.git/
Initialized SwarmGit project at /home/user/my-api
Run 'swarm task "your task"' to get started

Creates:

• swarm.yml — your orchestration config
• .swarm/ directory — where worktrees live
• .gitignore entry for .swarm/
• Initializes Git if not already a repo

---

swarm task "description"

Preview how a task will be split into subtasks before spawning agents.

$ swarm task "Build a REST API with JWT auth and user CRUD"
Analyzing task...

Proposed subtasks:
  1. Setup project structure and dependencies
  2. Implement JWT authentication
  3. Build user CRUD endpoints
  4. Write unit and integration tests

4 agents will be spawned.
Accept? [Y/n]: y

---

swarm run "description" [options]

All-in-one command: split, spawn, monitor, and optionally merge.

# Basic usage
$ swarm run "Build a todo REST API"

# Use a specific agent for all roles
$ swarm run "Build a todo REST API" --agent claudecode

# Auto-merge when agents finish
$ swarm run "Build a todo REST API" --auto-merge

# Assign different agents per role
$ swarm run "Build a todo REST API" --coder opencode --tester codex

# Limit parallel agents
$ swarm run "Build a todo REST API" --agents 2

---

swarm status [--watch] [--json]

Show current state of all agents.

# One-time status
$ swarm status
SwarmGit — my-project
──────────────────────────────────────────────────────────
AGENT      TOOL          BRANCH              STATUS        LAST COMMIT
agent-1    claudecode    swarm/agent-1       ✅ done       "setup project structure"
agent-2    opencode      swarm/agent-2       🟡 working    "added login endpoint"
agent-3    codex         swarm/agent-3       ⏳ waiting    —
agent-4    aider         swarm/agent-4       ⏳ waiting    —
──────────────────────────────────────────────────────────
1 done · 1 working · 2 waiting · 0 failed

# Real-time dashboard (refreshes every 2 seconds)
$ swarm status --watch

# Machine-readable output for scripting
$ swarm status --json
{
  "project": "my-project",
  "task": "Build a REST API with auth and tests",
  "agents": [
    {
      "name": "agent-1",
      "tool": "claudecode",
      "branch": "swarm/agent-1",
      "status": "done",
      "last_commit": "setup project structure"
    }
  ]
}

Status values:

• ⏳ waiting — not yet started
• 🟡 working — agent process running
• ✅ done — agent finished, ready to merge
• ❌ failed — agent process exited with error
• 🔀 merged — branch merged into main

---

swarm merge [options]

Merge finished agent branches into main.

# Merge all DONE agents
$ swarm merge
Merging agent-1 (swarm/agent-1) → main ...  ✅ clean
Merging agent-2 (swarm/agent-2) → main ...  ⚠️  conflict in src/user.js
  → Opening conflict resolution ...
  Press Enter when you have resolved and committed...

# Merge a specific agent
$ swarm merge --agent agent-2

# Auto-merge as agents finish (keeps running)
$ swarm merge --auto

# Preview what would be merged without doing it
$ swarm merge --dry-run
Dry run — would merge:
  agent-1 (swarm/agent-1)
  agent-2 (swarm/agent-2)

---

swarm peek <agent> [options]

Inspect a branch without switching to it.

# Show agent info
$ swarm peek agent-2
Agent: agent-2
Tool: opencode
Branch: swarm/agent-2
Worktree: /home/user/my-project/.swarm/agent-2
Status: done
Task: Implement JWT authentication

# Show file tree
$ swarm peek agent-2 --tree
.
├── src/
│   ├── auth.py
│   └── models.py
└── requirements.txt

# Read a specific file
$ swarm peek agent-2 --file src/auth.py
def authenticate_user(token: str) -> dict:
    ...

# Show git log for this branch
$ swarm peek agent-2 --log
a1b2c3d added jwt middleware
e4f5g6h implemented login endpoint

---

swarm diff <agent>

Show diff between an agent branch and main.

$ swarm diff agent-2
diff --git a/src/auth.py b/src/auth.py
new file mode 100644
index 0000000..abc1234
--- /dev/null
+++ b/src/auth.py
@@ -0,0 +1,45 @@
+import jwt
+from datetime import datetime, timedelta
+
+def create_token(user_id: str) -> str:
+    ...

---

swarm agent <subcommand>

Manual agent management.

# Create a custom agent manually
$ swarm agent create --name "auth-agent" --branch "feature/auth" --tool claudecode
Created agent auth-agent on branch feature/auth

# Assign a task
$ swarm agent assign "auth-agent" "implement JWT login and refresh token"
Assigned task to auth-agent

# List all agents
$ swarm agent list
agent-1: claudecode (done) — Setup project structure
agent-2: opencode (working) — Implement JWT authentication
agent-3: codex (waiting) — Build user CRUD endpoints

# Stop a running agent
$ swarm agent kill "agent-2"
Killed agent-2 (pid 12345)

---

swarm doctor

Check system requirements.

$ swarm doctor
SwarmGit Doctor
─────────────────────────────────────
✅ Git 2.43.0 (worktree supported)
✅ Python 3.11.4
✅ claudecode found at /usr/local/bin/claude
✅ opencode found at /usr/local/bin/opencode
⚠️  codex not found (install: npm install -g @openai/codex)
❌ aider not found (install: pip install aider-chat)
─────────────────────────────────────
2 agents ready · 1 warning · 1 missing

---

swarm clean [options]

Remove all worktrees and agent branches.

# Clean up everything
$ swarm clean
Removed worktree: /home/user/my-project/.swarm/agent-1
Deleted branch: swarm/agent-1
Removed worktree: /home/user/my-project/.swarm/agent-2
Deleted branch: swarm/agent-2
Cleaned up.

# Force remove even if agents are running
$ swarm clean --force

# Remove worktrees but keep branches
$ swarm clean --keep-branches

---

Configuration

SwarmGit is configured via swarm.yml at your project root.

project: my-project

splitter:
  provider: anthropic           # anthropic | openai | ollama
  model: claude-sonnet-4-5
  api_key_env: ANTHROPIC_API_KEY

max_agents: 4
worktree_dir: .swarm
branch_prefix: swarm/

auto_merge: false
merge_strategy: sequential
on_conflict: ask
on_done: notify

default_agent: claudecode

agents:
  architect: claudecode
  coder: opencode
  tester: codex
  reviewer: aider

agent_config:
  claudecode:
    model: claude-sonnet-4-5
    extra_flags: []
  aider:
    model: gpt-4o
    extra_flags:
      - --yes
      - --no-git

custom_agents:
  mytool:
    command: "mytool run {task}"
    done_signal: "Task complete."
    workdir_flag: "--cwd"
    env:
      API_KEY: "{MYTOOL_API_KEY}"

See swarm.yml.example for the full reference.

---

Architecture

┌─────────────────────────────────────────────────────────┐
│                        CLI Layer                        │
│   swarm init / run / status / merge / peek / agent      │
└───────────────────────┬─────────────────────────────────┘
                        │
┌───────────────────────▼─────────────────────────────────┐
│                    Orchestrator                         │
│   - Receives task string                                │
│   - Calls LLM to split into subtasks                    │
│   - Coordinates all modules below                       │
└──┬──────────────┬──────────────┬────────────────────────┘
   │              │              │
   ▼              ▼              ▼
┌──────────┐ ┌─────────┐ ┌──────────────┐
│ Worktree │ │  Agent  │ │    Merge     │
│ Manager  │ │ Monitor │ │ Coordinator  │
│          │ │         │ │              │
│ add      │ │ spawn   │ │ queue        │
│ remove   │ │ watch   │ │ merge        │
│ list     │ │ signal  │ │ conflict     │
└────┬─────┘ └────┬────┘ └──────┬───────┘
     │            │             │
     ▼            ▼             │
┌─────────────────────┐         │
│   Adapter Layer     │         │
│                     │         │
│  ClaudeCodeAdapter  │         │
│  OpenCodeAdapter    │         │
│  CodexAdapter       │         │
│  AiderAdapter       │         │
│  CustomAdapter      │         │
└─────────────────────┘         │
                                │
┌───────────────────────────────▼─┐
│         Git Layer               │
│   git worktree add/remove/list  │
│   git merge / git log           │
│   git status / git diff         │
└─────────────────────────────────┘

Core Modules

Module	Responsibility
swarmgit/cli.py	Click CLI entry point and argument parsing
swarmgit/orchestrator.py	Main coordination logic, state persistence
swarmgit/config.py	swarm.yml loader and validation
swarmgit/models.py	Data classes: AgentRun, Worktree, SwarmTask
swarmgit/git/worktree.py	Native git worktree subprocess wrappers
swarmgit/git/merge.py	Merge coordinator with conflict strategies
swarmgit/monitor/agent_monitor.py	Threaded agent process launcher and watcher
swarmgit/adapters/	Agent-specific command builders and done-detection
swarmgit/ui/status.py	Rich terminal dashboard

---

Agent Adapters

SwarmGit uses an adapter pattern to support any AI coding agent. Each adapter implements:

• build_command() — Construct the agent's CLI invocation
• is_done() — Detect when the agent has finished its task
• pre_run() / post_run() — Lifecycle hooks for setup and cleanup

Adding a New Agent Adapter

1. Create swarmgit/adapters/youragent.py
2. Extend BaseAdapter
3. Implement build_command() and is_done()
4. Register in swarmgit/adapters/__init__.py
5. Add tests in tests/test_adapters.py

See the Contributing Guide for details.

---

Contributing

We welcome contributions from the AI developer tools community!

git clone https://github.com/rjben/swarm-git.git
cd swarm-git
pip install -e ".[dev]"

# Run tests
pytest

# Format and lint
black swarmgit/
ruff check swarmgit/

# Type check
mypy swarmgit/

Good First Issues

• Add Ollama support for local LLM task splitting
• Add --follow support to swarm agent logs
• Build a Docker adapter for containerized agents
• Add swarm report to aggregate agent failures
• Implement parallel merge strategy (advanced)

---

Roadmap

Version	Features
v0.1.0 (MVP)	Core worktree orchestration, 4 built-in adapters, sequential merge
v0.2.0	LLM task splitting, swarm task preview, custom agents
v0.3.0	swarm peek, swarm diff, enhanced conflict resolution
v0.4.0	Parallel merge strategy, agent logs, failure reporting
v1.0.0	Stable API, plugin ecosystem, CI/CD integrations

---

License

MIT License — see LICENSE

Copyright (c) 2025 SwarmGit Contributors

---

Keywords

AI coding agents, multi-agent orchestration, git worktree, parallel coding, Claude Code, Aider, Codex, OpenCode, developer tools, CLI, Python, open source, Git workflow, AI-assisted development, agent swarm, branch management, merge automation, software development automation

---

Built with ❤️ for the AI coding community