guardian-mcp

Guardian MCP — Agent Skill

Guardian is a sidecar daemon for ML training. It monitors, detects anomalies, makes LLM-guided recovery decisions, and exposes everything through 35 MCP tools. Training scripts need zero changes.

---

Setup & Configuration

Prerequisites

• Python 3.10+
• pip install guarftrain (or clone + pip install -e .)
• MCP SDK: pip install guarftrain[mcp]
• For LLM-powered agent decisions: pip install guarftrain[agent] + set ANTHROPIC_API_KEY

Start Guardian MCP Server

Three modes, pick one based on your scenario:

A. Co-located with training (best real-time, recommended)

guarftrain watch --with-mcp -- python train.py --epochs 20

MCP server runs in the same process as guardian watchdog. Shares memory state directly.

B. Standalone stdio (Claude Code subprocess model)

guarftrain serve --transport stdio

Reads checkpoint/log files from disk. Suitable when training is already running separately.

C. One-click: Dashboard + MCP + browser

guarftrain start

Launches Dashboard (port 8765) + MCP SSE server (port 8766) + opens browser.

Connect Your Agent Client

Claude Code / Qoder — stdio mode:

Add to .claude/mcp.json or .claude/settings.local.json:

{
  "mcpServers": {
    "guardian": {
      "command": "guarftrain",
      "args": ["serve", "--transport", "stdio"],
      "cwd": "/path/to/your/project"
    }
  }
}

Remote server — SSE mode:

# On remote machine:
guarftrain serve --transport sse --port 8766

# SSH tunnel from local:
ssh -L 8766:127.0.0.1:8766 user@server

# Local client config:
{
  "mcpServers": {
    "guardian-remote": {
      "type": "http",
      "url": "http://127.0.0.1:8766/sse"
    }
  }
}

Enable Write Tools (optional)

Write tools (intervention, restart, stop) require explicit opt-in:

1. In configs/guardian.yaml:

mcp:
  enabled: true
  enable_write_tools: true

1. Set auth token:

export GUARDIAN_MCP_TOKEN=your-secret-token