🪖 ClaudeArmy

An MCP server that lets Claude orchestrate multiple Claude Code agents working autonomously across different projects.

What It Does

ClaudeArmy gives Claude (in the chat interface) the ability to spawn background Claude Code processes that work independently on coding tasks. Think of it as a chain of command:

• You → give strategic direction
• Claude (chat) → breaks it down and dispatches agents
• Claude Code agents → execute autonomously, spawning their own sub-agents as needed

This means you can kick off work across multiple projects simultaneously, keep chatting normally, and check in on progress whenever you want.

Tools

Setup

1. Clone this repo
2. npm install
3. Add to your Claude Desktop config (~/.config/Claude/claude_desktop_config.json):

"claude-army": {
  "command": "node",
  "args": ["/path/to/claude-army/src/index.js"]
}

4. Restart Claude Desktop

Usage Examples

Single project task:

"Refactor GarbageFire's app.py into modular components"

Multi-project parallel work:

"Refactor GarbageFire to be more modular, and add a Salesforce adapter to my Reporting Tool"

Claude will dispatch separate agents to each project directory and let them work autonomously.

Check progress:

"How are the agents doing?"

Configuration

Edit the constants at the top of src/index.js:

• CLAUDE_BINARY - Path to your Claude Code binary
• MAX_CONCURRENT_TASKS - Max simultaneous agents (default: 5)

Requirements

• Claude Code installed and authenticated
• Node.js 18+
• Claude Desktop or any MCP-compatible client

Changelog

v0.3.0

• Fixed stream-json parsing to correctly handle assistant message content arrays
• Staleness indicator: check_tasks now shows time since last agent activity
• Better error surfacing: actionable messages for CLI not found, permission denied, and mid-run failures
• Fixed output duplication in get_task_output
• Demo GIF added to README

v0.2.0

• Real-time progress tracking via Claude Code's stream-json output
• check_tasks now shows recent agent activity (file reads, edits, bash commands) instead of just runtime
• get_task_output includes a full progress timeline with timestamps
• Structured event parsing for tool use, assistant messages, and results

v0.1.0

• Initial release
• Core tools: dispatch_task, check_tasks, get_task_output, cancel_task, purge_tasks
• Fire-and-forget architecture with background process management
• Max 5 concurrent agents
• stdio transport for MCP communication

License

MIT