The shared layer for public data
Hundreds of government datasets, cleaned and unified into one queryable API. Point at a question, get an answer back. No format hunting, no download queues, no crosswalk spreadsheets.
The problem
There are over 300,000 public datasets on data.gov alone. Teacher salaries, poverty rates, workplace injuries, census demographics. All technically "open." Try to actually use one.
You'll find ZIP files inside ZIP files. Excel spreadsheets where the columns shift every year. Documentation pages that 404. APIs that haven't been maintained since 2019. And when you finally get the data into a usable format, you'll discover that the Census Bureau and the Bureau of Labor Statistics use different geographic identifiers, so the join you need doesn't exist yet. You have to build the crosswalk yourself.
This is the part nobody talks about. The data is public. The infrastructure to make it usable isn't. Researchers spend roughly half their time cleaning and wrangling before any actual analysis begins. The bottleneck isn't access. It's the gap between "technically available" and "actually queryable."
What it actually does
OpenData automates the work nobody wants to do. Government CSVs, Excel files, and APIs get fetched, parsed, typed, and normalized into Parquet files. From there, they're queryable via DuckDB SQL through a single API endpoint. Same syntax, same response shape, regardless of where the data came from.
That means a 100-million-row dataset that would choke a spreadsheet and overwhelm an LLM's context window becomes something you can write a precise SQL query against. Filter by state and year. Aggregate by agency. Join Census income data with BLS wage data on the county code that already exists in both. You get exactly what you asked for, not the whole file.
A knowledge graph layer pre-computes which datasets share joinable columns: geographic identifiers, time series keys, common codes. Cross-dataset joins that used to require a week of research to even discover are surfaced automatically.
For AI agents, this architecture matters a lot. When Claude, ChatGPT, or Cursor connects via the MCP server, it can search for relevant datasets, inspect the schema, and write a SQL query that returns exactly what a user asked for. Instead of trying to read a 50MB CSV in full, the agent gets structured metadata and runs a precise query. The data stays current through automated ingestion pipelines that sync from original government sources. No manual uploads, no contributor queues.
Get started with the quickstart guide, or browse the catalog.
How this started
Riley Hilliard. Austin, TX. Software engineer.
This project started with a dinner table argument. My dad bought a duplex as an investment while he was in college. I couldn't have done the same, and he didn't believe me. Proving it meant gathering housing prices, income data, inflation rates, and cost-of-living breakdowns from scattered government portals, each with its own format and quirks. What should have taken an afternoon took weeks.
That friction is what OpenData exists to eliminate. The data already exists. Making it usable shouldn't be a research project in itself.
The full story is on Labs: 28 Years, 14 Formats, One API Call and The American Dream Now Costs Three Times What It Used To.
Open source
The core of OpenData, the ingestion engine, query layer, connectors, CLI, and API, is Apache 2.0 licensed. The MCP server is open source too. The web platform and cloud features (search, user accounts, collaboration) are proprietary.
This split is intentional. The infrastructure should be inspectable, forkable, and verifiable by anyone who wants to run it. The hosted platform captures enough value to fund continued development. You can use the open source library standalone without ever touching the cloud product, and the cloud product is just the library running at scale with a frontend on top.
Read more about the open core model, or see the source on GitHub.