Digible Data — an MCP server

Six composable tools over a multifamily marketing dataset. Ask about money out, people through the leasing funnel, and phone calls — grouped however you need.

Live: https://mcp-digible-queries.fastmcp.app/mcp — public, no sign-in, no key.

The tools

Three shapes — money, people, phone — plus three lookups. The variety lives in group_by, not in more tools. leasing_funnel(include_cost=True) joins spend on the same grouping key, so cost per lead and cost per lease need no extra tool.

period takes what a person would say: "May", "Q2", "last month", "first half", or "2026-02-01..2026-04-15". If it cannot read one, it says so and lists what it does understand.

Every tool takes an optional caller for attribution. It defaults, so nobody using Claude Desktop has to invent a name.

Related MCP server: Northbeam MCP Server

Connecting it

Claude Desktop — add it as a connector with the URL above, then ask in plain language. Nothing to install.

Claude Code

claude mcp add --transport http digible-data https://mcp-digible-queries.fastmcp.app/mcp

Locally

pip install -r requirements.txt
python server.py            # http://127.0.0.1:8030/mcp
python server.py --stdio    # or stdio, if your client prefers it

Where this came from

Built for the Overclock Accelerator MCP training as the worked answer to an exercise: six agents, each with its own way of querying the same database, consolidated into one server. Agent code dropped from 942 lines to 141.

The exercise and the write-up live in digible-mcp-training-starter.

Data

digible.db is synthetic, committed, and opened read-only. Properties, campaigns, channels, spend, leads, calls, tours, applications and leases, covering 2026-01 to 2026-06.

MIT licensed.