AI-native architecture

Built for agents. Usable by humans.

Jeevesy is a tenant operating system where MCP tools are the primary API, LLMs are a runtime primitive, and the dashboard is a thin client over the same tools your agents call.

1. Tool layer (MCP)

Every capability is an MCP tool

Cases, bookings, rent, contracts, journeys, documents, communications — typed schemas, RLS-scoped, versioned. The tenant app, admin dashboard and external agents all call the same tools.

2. Reasoning layer

LLMs as a runtime primitive

Triage, categorisation, sentiment, translation, drafting and journey decisions run through models by default. Prompts are versioned and evaluated like code.

3. Agent layer

Jeeves and your own agents

Jeeves is the built-in resident/operator concierge. Bring ChatGPT, Claude, Cursor or an in-house agent — they connect to the same MCP endpoint with per-user OAuth.

4. Interop layer

MCP in, MCP out

Register external MCP servers (work-order systems, calendars, energy providers) and Jeevesy's agents call them alongside native tools. No brittle integrations.

5. Identity & safety

Per-user auth, end-to-end

Supabase OAuth 2.1 with dynamic client registration. Every tool call runs as the authenticated user, enforced by database-level RLS. Mutating tools require explicit approval.

6. UI layer (fallback)

A dashboard for the humans

The web app is a thin client over the tool layer — every screen is generated from the same schema an agent sees. Click when you prefer clicking; prompt when you don't.