SWAT

Autonomous multi-agent task execution engine. SWAT dispatches tasks to specialist squads — each running as an independent AI coding agent session.

🚀 Installation

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/LangSensei/swat/main/install.sh | bash

Windows

irm https://raw.githubusercontent.com/LangSensei/swat/main/install.ps1 | iex

This will:

1. Download the latest release for your platform
2. Install the SWAT binary to ~/.swat/bin/ and add it to your PATH (Linux/macOS: appends to .bashrc/.zshrc; Windows: updates user PATH in the registry)
3. Install framework blueprints to ~/.swat/blueprints/

OpenClaw Integration (Optional)

If you use OpenClaw, install the bridge plugin separately:

# See https://github.com/LangSensei/swat-openclaw

🗑️ Uninstallation

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/LangSensei/swat/main/uninstall.sh | bash

Windows

irm https://raw.githubusercontent.com/LangSensei/swat/main/uninstall.ps1 | iex

Add --purge to also remove runtime data (operation history).

🎮 Usage

SWAT exposes 12 MCP tools. Configure your agent to connect:

MCP Configuration

{
  "mcpServers": {
    "swat": {
      "command": "swat",
      "args": []
    }
  }
}

Tools

Operations

• swat_dispatch — Dispatch a task (auto-classified to the right squad)
• swat_ops — List operations with filters (status/since/limit/offset)
• swat_cancel — Cancel a running operation

Squads

• swat_squads — List installed squads
• swat_squad_browse — Browse the marketplace
• swat_squad_install — Install a squad from the marketplace
• swat_squad_uninstall — Uninstall a squad
• swat_squad_update — Update a squad to the latest marketplace version

Intake

• swat_intake_create — Create a recurring task (zero LLM cost)
• swat_intake_list — List all intake queue entries
• swat_intake_delete — Delete an intake entry

Notification

• swat_notify — Send a notification to the user

CLI Flags

Flag	Description
--version	Print the installed version and exit
--mcp-only	Start only the MCP server — skip the background commander loop
--runtime <name>	Set the AI coding agent runtime (default: copilot)
--notify <target>	Set the notification target: desktop, openclaw (default: desktop)

🧠 Architecture

Agent → MCP → SWAT Commander (Go) → Squads (AI coding agent sessions)

SWAT is runtime-agnostic. The Commander orchestrates operations while a pluggable RuntimeAdapter handles runtime-specific details (dot-directories, agent files, MCP config paths, hooks).

Component	Role
Commander	Go MCP server. Handles dispatch, workspace composition, dependency resolution, scheduling, and completion scanning.
Squads	Specialist agents. Each runs as an independent coding agent session with its own skills, MCP tools, and protocol.
RuntimeAdapter	Abstracts runtime differences. Each adapter knows its dot-dir, agent file name, MCP config path, and how to launch the agent.

Supported Runtimes

Runtime	Dot-Dir	Agent File	MCP Config	Launch Command
copilot (default)	.github/	AGENTS.md	.mcp.json	copilot
gemini	.gemini/	GEMINI.md	.gemini/settings.json	gemini

Set the runtime with --runtime <name> (e.g. swat --runtime gemini).

How It Works

1. You dispatch a task (Commander auto-classifies to the right squad)
2. Commander provisions the workspace — copies squad snapshot to .squad/, skill hooks to <dotdir>/hooks/, skill content to <dotdir>/skills/, resolves MCP dependencies, writes the agent file (runtime-specific)
3. A coding agent session launches in the operation directory, reads the agent file + OPERATION.md
4. Commander's background loop scans for completion (OPERATION.md status + report.html)
5. Results surface through your agent

Skills vs Hooks: Skills are runtime-agnostic (shared across all runtimes). Hooks are runtime-specific and live under hooks/copilot/, hooks/gemini/, etc. in each skill's blueprint.

Intake Queue

Unified task entry point — all tasks (immediate dispatches and recurring schedules) flow through ~/.swat/intake/:

• Immediate tasks: created by swat_dispatch, processed by background loop, deleted after completion
• Recurring tasks: standard 5-field cron expressions with timezone support
• immediate flag for first-run-now behavior
• Per-file locking for concurrent safety

📂 Directory Structure

~/.swat/
├── blueprints/                    # Templates & installed content
│   ├── OPERATION.md               # Operation template
│   ├── squads/
│   │   ├── _framework/            # Core protocol (versioned with binary)
│   │   │   ├── PROTOCOL.md
│   │   │   └── TEMPLATE.md
│   │   └── <squad>/               # Squad blueprints
│   │       └── MANIFEST.md
│   ├── skills/                    # Shared skills (runtime-agnostic)
│   └── mcps/                      # MCP server configs
│
├── squads/                        # Runtime data
│   ├── _unclassified/             # Operations before squad assignment
│   └── <squad>/
│       └── operations/
│           └── <id>/              # Operation workspace
│               ├── OPERATION.md
│               ├── <agent-file>   # Runtime-specific (AGENTS.md or GEMINI.md)
│               ├── <mcp-config>   # Runtime-specific (.mcp.json or .gemini/settings.json)
│               ├── report.html
│               ├── .squad/        # Squad blueprint snapshot (read-only)
│               └── <dotdir>/      # Runtime-specific (.github/ or .gemini/)
│                   ├── hooks/     # Skill hooks (runtime-specific)
│                   └── skills/    # Skill content
│
└── intake/                        # Intake queue (JSON)
    └── <id>.json

🛠️ Prerequisites

• Linux, macOS, or Windows
• GitHub CLI (authenticated)
• At least one supported AI coding agent runtime:

Runtime	Requirements
Copilot (default)	GitHub Copilot CLI (npm install -g @github/copilot), Node.js 18+
Gemini	Gemini CLI installed and available as gemini on PATH

🔨 Building from Source

go build -o swat .   # Go 1.24+
# With version injection:
go build -ldflags "-X main.version=v1.0.0" -o swat .

📦 Related Projects

• swat-marketplace — Squads, skills, and MCPs
• swat-openclaw — OpenClaw integration (plugin + skill)

📄 License

MIT