The landing page is the pitch. This page is the project, in full.
Large language models learn from, and search over, the internet: a web of pages written for human readers. It is a magnificent archive of information, and a terrible source of data. Numbers on the web are quoted, rounded, outdated, stripped of context, and copied from copies. When an AI system answers a question about inflation, housing, employment, or trade, it is usually reasoning over prose that mentions numbers rather than the numbers themselves. That is a large part of why confident, wrong figures appear in AI answers.
Meanwhile the actual record exists. Statistical agencies and public institutions publish the measured state of the world: prices, populations, permits, payrolls, output, trade, in the thousands of tables, continuously revised and maintained. But that record lives in hundreds of inconsistent portals, formats, and APIs, built for human analysts with time to spare. There is an internet of information. There has never been an internet of data.
The vision of a machine-readable web of structured data is not new. It was proposed in 2001 as the Semantic Web, and it consumed two decades of standards work without taking off, for three reasons: publishers had no incentive to structure their data, no software existed that could consume messy heterogeneous data, and the standards were too heavy to adopt.
Every one of those failure conditions has now reversed. AI agents are the consumers the Semantic Web never had, and they tolerate imperfect data. The Model Context Protocol is the lightweight interface its standards never became. And modern AI collapses the cost of building the connectors themselves. Starwell is that old idea, rebuilt for the moment it became possible.
Starwell is a verified data layer for AI: the world's official statistics, ingested into one harmonized store and served through one API and one MCP server. Four capabilities sit on top of the store:
/data returns verified series in a single harmonized schema, with full provenance on every observation and complete revision history, so "what did this number say on March 3rd" is always answerable./answer takes a plain-language question and returns a computed answer, the chart, the Python that produced it, and the citation to the official source table./monitor watches series and alerts when they change or break trend.Every connector must pass a conformance suite and a golden-value audit: known, published reference figures are pulled through the connector and matched against the official source before any series earns a verified status. Checks re-run on schedule to catch upstream changes. Nothing is served without a verification status, and nothing is stored without provenance.
Two things we will not do. We will never claim to eliminate hallucinations; our claim is narrower and provable: answers computed from verified official data, cited to the source table. And we will never serve data we cannot cite openly. Only openly licensed official sources enter Starwell, each carrying its license and attribution. If we cannot print the citation proudly, it does not go in the well.
The first stars: Statistics Canada, FRED, and the U.S. Bureau of Labor Statistics, built to the highest standard rather than the fastest schedule. SEC EDGAR and the World Bank followed, extending the well to company filings and to more than two hundred countries. From there, an agent-assisted connector pipeline scales coverage across national and regional statistical portals, with every new connector passing the same verification gate, and a community program that lets contributors add sources and share in the revenue their connectors earn.
And above the layer sits a research program. Once the well is deep enough, the corpus itself becomes something no one has assembled before: the planet's official record, harmonized, provenance-clean, and revision-aware. Training on it is the mission of Earthlight.
Starwell is built by Datastam (Blue Green Yellow Inc.), a Toronto company whose analysis engine already powers automated data analysis for public data at canada.datastam.ai. Starwell is deliberately built lean: a small human core directing an agent workforce that discovers portals, builds connectors, audits every value, and monitors freshness, with a human approving every merge.
General and partnership inquiries: hello@datastam.com
Investors: Starwell is infrastructure for the agent economy with a long-horizon research program on top. If that is the kind of boundary you fund, reach us through the contact form or at hello@datastam.com.
Developers: join the waitlist and tell us what your agent needs first.