AI and MCP

#AI-assisted risk

Under Settings > AI, point CroustyPatch at any OpenAI-compatible chat API (Ollama, vLLM, LocalAI, OpenAI…): base URL, model and an optional key (stored encrypted). Test provider makes a real round trip.

A policy with AI mode asks the model for one thing only: the contextual risk of a batch. Stability, the decision table, waivers, windows, approvals, soak and rollback all stay deterministic.

Guardrails:

• the answer must be valid JSON with a known level, or it is ignored;
• the model may raise the risk freely but lower it by one level at most;
• any error, timeout or bad answer falls back to the deterministic engine, and the task says so;
• only host attributes, package names, versions and CVE metadata are sent, never keys or logs.

On any task, Ask the AI gives an advisory second opinion that changes nothing and is audited. Small models answer in the right format but reason poorly: use a 7B model or larger for real decisions, and keep an unauthenticated Ollama on a management network.

#MCP server

POST /api/v1/mcp speaks the Model Context Protocol (JSON-RPC 2.0). Authenticate with a personal API token (My account > API tokens, Authorization: Bearer cpat_…). It acts as you, with your role and permissions, and shows in the audit trail. Tokens are stored hashed, shown once, expire in 1 to 365 days and can be revoked.

Tools: fleet_summary, list_hosts, get_host, list_tasks, get_task (read-only) and request_patch, which always creates a task awaiting human approval. An assistant never installs anything by itself.

{ "mcpServers": { "croustypatch": { "type": "http", "url": "https://patch.example.org/api/v1/mcp",
  "headers": { "Authorization": "Bearer cpat_…" } } } }