mcp-supervisor

MCP server that exposes runtime-supervisor as tools for LLM agents (Claude Desktop, Claude Code, any MCP-aware client).

Install

uv pip install -e packages/mcp-supervisor
# or from PyPI once published:
pip install mcp-supervisor

Configure

Environment variables:

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "runtime-supervisor": {
      "command": "uv",
      "args": ["--directory", "/path/to/agentic-internal-controls", "run", "mcp-supervisor"],
      "env": {
        "SUPERVISOR_URL": "http://localhost:8000",
        "SUPERVISOR_APP_ID": "your-integration-id",
        "SUPERVISOR_SECRET": "your-shared-secret",
        "SUPERVISOR_SCOPES": "refund,payment"
      }
    }
  }
}

Exposed tools

• evaluate_action(action_type, payload, dry_run=false) — gate an agent action through policy + risk.
• list_action_types() — discover what's supervised (live vs planned).
• get_evidence(action_id) — fetch the hash-chained evidence bundle for audit.

Example prompt to Claude

"Before I execute a $1200 refund for a customer who signed up 18 days ago and has had 2 refunds in the last 24h, use evaluate_action with dry_run=true to show me what the supervisor would decide."