belle-mcp-server

A reference Model Context Protocol (MCP) server that exposes real Belle Realty property-management data — properties, tenants, leases, maintenance tickets, rent roll — as tools that Claude Desktop, Cursor, or any MCP-compatible client can call directly.

Six tools. Five are strictly read-only. One is a HITL-gated propose-write. That ratio is intentional and is the whole point of this repo.

Part of the AI Fluency Program — Level 2.

Why this exists

Most "AI + your data" demos give the model unrestricted database access. That is a footgun.

The Model Context Protocol is designed to expose a small, curated surface area with per-tool auth, rate limits, and audit — the same discipline you'd apply to a public REST API. This repo shows what that looks like for a real domain (a Louisiana shopping center) with a real Postgres schema, a working seed, and a single HITL-gated write path.

If you understand this repo, you can build one for any business you run.

What you get

Every call is rate-limited (60/min default) and audit-logged to mcp_audit_log.

Quick start

# 1. Clone + install
git clone https://github.com/OrangeOnyx/belle-mcp-server.git
cd belle-mcp-server
npm install

# 2. Configure
cp .env.example .env
# Paste your Supabase URL + service-role key

# 3. Set up the schema (Supabase project)
#    Copy supabase/migrations/0001_init.sql into the SQL editor and run.

# 4. Seed demo data
npm run db:seed

# 5. Build + inspect
npm run build
npm run inspect

The MCP Inspector opens a UI where you can list tools, call them, and see raw responses.

Wire it into Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on Windows/Linux:

{
  "mcpServers": {
    "belle-realty": {
      "command": "node",
      "args": ["/absolute/path/to/belle-mcp-server/dist/index.js"],
      "env": {
        "SUPABASE_URL": "https://your-project.supabase.co",
        "SUPABASE_SERVICE_ROLE_KEY": "your-service-role-key"
      }
    }
  }
}

Restart Claude Desktop. You'll now see a belle-realty toolset. Try:

"What suites are currently occupied at On The Boulevard, and how much monthly rent are they producing?"

Claude will call get_rent_roll and answer from the returned data.

The HITL write pattern

The one write tool (draft_maintenance_response) illustrates a general pattern you should copy for any AI-facing service:

1. AI proposes a change — here, a reply to a tenant maintenance ticket.

2. The server saves it as approved=false.

3. Nothing is delivered, sent, or applied until a human approves it out-of-band (typically in the property manager's admin UI).

4. The MCP surface intentionally does not expose an approve tool. Approval is a human-only operation.

This means an over-eager or prompt-injected agent cannot silently push text to a tenant. It can propose, and it can propose loudly. It cannot ship.

For a longer walkthrough, see docs/hitl-pattern.md.

Personal-use walkthrough

You're an individual landlord with 3 rental houses or one small commercial building.

1. Run the migration on your Supabase project.

2. Seed with your own data (edit supabase/seed.ts, or insert rows manually).

3. Point Claude Desktop at the server.

4. Ask questions like "which tenant has a lease expiring in the next 90 days?" or "draft a response to the ticket about the water heater."

You've now built an AI-native tenant operations layer that speaks your data. It cost you an evening.

Company-use walkthrough

You run Belle Realty (or an equivalent management company). Multiple staff need Claude access to portfolio data without seeing raw SQL, and without any risk of accidental writes.

1. Deploy this server as a persistent process (Railway, Fly, or a Docker host).

2. Set MCP_TRANSPORT=http and MCP_HTTP_TOKEN=<shared-secret>.

3. Each teammate configures Claude Desktop or Cursor with the URL + token.

4. Read-only tools give everyone leverage. The single write tool protects the tenant relationship.

5. mcp_audit_log gives you an after-the-fact record of every AI action.

Architecture

graph LR
    A[Claude Desktop / Cursor] -->|MCP stdio or HTTP| B[belle-mcp-server]
    B --> C[RateLimiter]
    B --> D[Zod validation]
    B --> E[Supabase Postgres]
    B --> F[mcp_audit_log]
    E --> G[(properties, tenants, leases, tickets)]

Details in docs/architecture.md.

Extending it

Add a new tool in 4 steps:

1. Add a Zod schema for input in src/schemas/domain.ts (if the data shape is new).

2. Create src/tools/<name>.ts with an input schema, a handler, and a JSON-Schema definition.

3. Register it in src/tools/index.ts.

4. Add tests in tests/.

Every write tool should follow the propose-write pattern in draft_maintenance_response.

Deploy

Railway (recommended for HTTP transport)

railway up

railway.json builds the server and runs node dist/index.js. Set env vars in the Railway dashboard.

Local (stdio only)

Just build and point your MCP client at dist/index.js. No hosting needed.

Development

npm run dev       # tsx watch mode
npm run test      # vitest
npm run build     # tsc → dist/
npm run inspect   # MCP Inspector UI

Related repos

• lease-abstractor — extract a structured abstraction from a lease PDF/DOCX

• support-triage-agent — same HITL pattern applied to support messages

• diligence-agent — RAG-based diligence over a document folder

• ai-fluency-program — parent curriculum

License

MIT — see LICENSE.

Not legal, tax, or property-management advice. Do not use for compliance-critical decisions without a licensed professional in the loop.