census-mcp

U.S. Census data by ZIP code, inside your agent. An MCP server that lets an LLM look up income, demographics, housing, and education for any ZIP — built on Anthropic's official mcp Python SDK.

All tools are read-only. Data is the U.S. Census Bureau's American Community Survey (ACS) 5-year release — bulk-downloaded once into a local store, then served offline. A free Census API key is required for the one-time download.

Status: published (v0.1.2) — uvx mcpwright-census (PyPI) and listed in the official MCP Registry as io.github.mcpwright/census-mcp. All 8 tools below are live and unit-tested, verified end-to-end against the live ACS 5-year release (vintage 2024 at the time of writing — the server auto-detects the latest). Part of the mcpwright suite.

Tools

Tool	What it does
`lookup_zip(zip_code)`	Confirm a ZIP maps to a Census ZCTA; returns name, total population, and the ACS vintage.
`get_income(zip_code)`	Median household income, per-capita income, total households, and the % of households earning $200k+.
`get_demographics(zip_code)`	Total population and median age.
`get_housing(zip_code)`	Median home value, median gross rent, occupied units, and % owner-occupied.
`get_education(zip_code)`	Of adults 25+, the % with a bachelor's degree or higher and the % with a graduate/professional degree.
`compare_zips(zips, metric)`	Rank several ZIPs by one metric (income, age, home value, attainment, …), highest first.
`get_acs_variable(zip_code, variable)`	Escape hatch: the raw value of any stored ACS variable, by code or friendly name.
`find_zips(place, state=None)`	Reverse lookup: the ZIPs that fall within a city/town/CDP, ranked by how much of each ZIP's land lies in the place.

ZIP ≈ ZCTA (ZIP Code Tabulation Area): they mostly coincide, but ~2% of ZIPs (PO-box-only / non-residential) have no ZCTA and will return an error. All ACS figures are 5-year estimates.

find_zips is the reverse direction (place → ZIPs). It's approximate: ZCTAs don't nest inside places, so a ZIP can span several places and vice versa — read each match's coverage_pct and pass a state to disambiguate same-named places (e.g. Cambridge, MA vs OH). It's built on the public 2020 Census ZCTA-to-Place relationship file (2020 geography; no API key needed for it).

Install

Requires Python 3.12+ and a free Census API key (set CENSUS_API_KEY). The PyPI package is mcpwright-census; the command, server, and tools are all "census".

server.ts

export CENSUS_API_KEY=your-free-key
uvx mcpwright-census setup     # one-time: bulk-download the ACS dataset into a local store

Claude Code

Terminal

claude mcp add census -e CENSUS_API_KEY=your-free-key -- uvx mcpwright-census

Claude Desktop

Add to claude_desktop_config.json:

config.json

{
  "mcpServers": {
    "census": {
      "command": "uvx",
      "args": ["mcpwright-census"],
      "env": { "CENSUS_API_KEY": "your-free-key" }
    }
  }
}

First run downloads the full ACS dataset (~33k ZCTAs) into a local SQLite store under your OS cache dir, so every later lookup is instant and offline. Run mcpwright-census setup ahead of time, or let the server download lazily on first use. mcpwright-census refresh re-pulls when a new ACS vintage is published. Override the store location with CENSUS_MCP_STORE.

Develop

bash

git clone https://github.com/mcpwright/census-mcp && cd census-mcp
uv sync
uv run pytest -v                                     # tests (mocked Census + seeded SQLite)
uv run ruff check src/ && uv run ruff format --check src/   # lint + format
uv run mypy                                          # type checking
uv run mcp dev src/census_mcp/server.py              # poke the tools in the MCP Inspector

Roadmap

Privacy

census-mcp runs entirely on your machine and collects, stores, or transmits no personal data — no accounts, no tracking, no telemetry. Its only outbound requests go to the public U.S. Census Bureau to download public data: api.census.gov for the ACS dataset (using your own CENSUS_API_KEY), and www2.census.gov for the place-lookup file used by find_zips (no key needed). The key is read from your environment and is never logged or sent anywhere else. These downloads happen during setup/refresh (or lazily on first use); after that, lookups are served from a local store with no further network calls. The downloaded data lives under your OS cache directory (public reference data only) and can be deleted at any time.

Full policy: https://mcpwright.com/privacy

Questions & feedback

Questions, ideas, or "could it do X?" → Discussions
Bugs & concrete feature requests → Issues

Contributions welcome.

Part of mcpwright · built by Devender Gollapally