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

🧠 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).

Supported Runtimes

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:

🔨 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