1. The 8 Narrow MCP Tools (2025-11-25)
UnitVO never exposes raw database mutation tools or generic SQL/HTTP tools to an AI model. The AI Service Assistant operates exclusively through 8 narrow, schema-validated tools testable live in the MCP & Verification Lab:
| Tool Name | Scope | Purpose |
|---|---|---|
context.get_service_context | Read-Only | Returns verified unit/room metadata, operating mode, SLA targets, and emergency numbers. |
knowledge.search_approved | Read-Only | Searches approved building SOPs, Wi-Fi/parking guides, and amenity hours. |
operations.create_action_item | Idempotent Write | Creates a validated maintenance, housekeeping, or front-desk action item. |
operations.get_my_request | Read-Only | Fetches live status, assigned technician, and ETA for the caller's request. |
operations.set_access_window | Idempotent Write | Sets or updates the resident's authorized unit entry window and presence rule. |
operations.request_human | Idempotent Write | Escalates the conversation to a human duty manager with structured context. |
booking.get_quote | Read-Only | Checks live CapacityLedger availability and returns a minor-unit price quote. |
booking.prepare_change | Idempotent Write | Places an atomic hold or prepares a stay extension/date change for confirmation. |
2. Strict Authority-Field Rejection
Every tool schema enforces additionalProperties: false and explicitly rejects caller-supplied authority fields:
Rejected Authority Fields:
- organization_id
- worker_id
- approved_amount
- priority_override
Organization, property, and actor identity are injected exclusively from the authenticated session token.
3. Running the 34 Reference Verification Tests
Click Run All 34 Reference Tests in MCP Lab or call POST /v1/reference-tests/run to execute all 34 synthetic tests covering worker eligibility, offer expiration, idempotency conflicts, last-room race conditions, multi-night rollback, and authority rejections.