HDB Price MCP

2025. A weekend project to explore MCPs (Model Context Protocol) — built something useful in the process.

the idea

MCP lets AI assistants call external tools directly. I wanted to understand how it worked, and I wanted to build something with real data. Singapore's HDB resale transaction records are public, go back to 1990, and cover every flat sale in the country. That's a good dataset to query.

The question: could I ask an AI assistant "what's the median price of a 4-room flat in Tampines?" and get a real answer backed by actual transaction data?

what was built

Three repos, three layers:

hdb-price-db — the data pipeline. Downloaded raw HDB resale CSVs from data.gov.sg covering 1990–2025, consolidated five separate files (different date ranges, different schemas) into a single clean dataset using Jupyter notebooks, then loaded it into Supabase.

hdb-price-api — a FastAPI backend deployed on Render.com. Single endpoint: query by town, flat type, flat model, and date range. Returns average, min, max, p50, and p90 prices plus transaction count. Backed by a Supabase RPC function for efficient aggregation.

hdb-price-mcp — wraps the API as an MCP server. Registered as a tool so AI assistants (Claude, Cursor, etc.) can call it mid-conversation to fetch real price data on demand.

stack

• Python, FastAPI, Uvicorn
• Supabase (PostgreSQL + RPC functions)
• Render.com (API hosting)
• Jupyter (data transformation)
• MCP SDK

what i learned

MCP is a clean abstraction once you get past the setup. The interesting part isn't the protocol — it's that the quality of your tool depends almost entirely on the quality of your data and your API design. A well-structured endpoint with good filtering makes the AI's answers dramatically better.

The HDB dataset itself is fascinating. 35 years of flat transactions across every town in Singapore. Price trends, estate maturity effects, flat type premiums — all queryable in a few seconds.