POST /api/v1/ai/intent takes natural language and returns structured intent: - 'Deploy Plex' → { intent: deploy, appId: plex, deployPlan } - 'I want to stream movies' → { intent: recommend, categories: [media-streaming] } - 'Why is Plex down?' → { intent: diagnose, serviceId: plex } - 'Back up everything' → { intent: backup } - 'Is everything OK?' → { intent: health } GET /api/v1/ai/capabilities returns self-describing capabilities for agent discovery. Pattern-based matching works offline (no LLM call needed). LLM_PROXY_URL env var can be set for complex query delegation. 18 intent tests covering deploy, recommend, diagnose, backup, health, list, and unknown intents. 1770 total tests pass.