mcp-supervisor

# mcp-supervisor MCP server that exposes runtime-supervisor as tools for LLM agents (Claude Desktop, Claude Code, any MCP-aware client). ## Install ```bash uv pip install -e packages/mcp-supervisor # or from PyPI once published: pip install mcp-supervisor ``` ## Configure Environment variables: | Variable | Required | Description | | --------------------- | -------- | ------------------------------------------------------ | | `SUPERVISOR_URL` | yes | Supervisor API base URL (e.g. `http://localhost:8000`) | | `SUPERVISOR_APP_ID` | yes | Integration id issued by `POST /v1/integrations` | | `SUPERVISOR_SECRET` | yes | Shared secret returned at registration | | `SUPERVISOR_SCOPES` | no | Comma-separated scopes (default `*`) | ## Claude Desktop Add to `claude_desktop_config.json`: ```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."