The Foursquare H3 Hub
Xun Li, a Software Engineer at Foursquare who works on the Foursquare Spatial Agent and the Foursquare H3 Hub, presented three free tools Foursquare has built on H3. The key point: when every dataset already shares the same H3 cells, an AI agent can answer a spatial question from a single sentence, because it only has to choose and weight layers rather than solve geometry. The talk was pre-recorded and played at the H3 Conference 2026 in New York City.
The H3 Hub: 103 datasets you query in place
The Foursquare H3 Hub hosts 103 open datasets, all H3-indexed. Most are at resolution 8, and some are published natively at resolution 7. They span eight categories, from boundaries and demographics to environment and reachability, and all of them are free.
The Hub is an Apache Iceberg catalog, so you point DuckDB, Spark or PyIceberg at it and query the tables directly. As Xun put it: no downloads, no copies. The catalog is not only static snapshots. It also includes time series published annually, monthly and even weekly.
Spatial Desktop: a map with a SQL engine behind it
Foursquare Spatial Desktop is a free, cross-platform app. It uses kepler.gl to render the map and DuckDB for compute on large datasets, so, as Xun noted, it is not just a viewer. You can run SQL against H3 Hub datasets or against your own data, and the app is optimized around the Hub's datasets.
An agent that only has to pick layers
The Foursquare Spatial Agent is the piece Xun said the team is most excited about. You ask a question in plain text, and the agent:
- searches the H3 Hub for datasets that could answer it
- tells you which ones it would pick and asks you to confirm
- queries the data where it lives
- filters to the boundary you mean
- composes the layers into an answer
Xun's explanation for why this works so well on H3 is that the geometry problem is already solved. Every layer is keyed on the same cells, so the agent only has to reason about which layers to use, which variables to pick and how to weight them. That, Xun said, is a much smaller problem. Foursquare is also bringing the same agent into the H3 Hub website, which was in staging at the time of the recording.
Three demos from a single sentence
Vineyard suitability in Italy. Xun asked the agent which areas of Italy have the best soil, climate and terrain for wineries. It proposed soil properties, climate normals, elevation and land cover. After Xun confirmed, it combined nine variables into a single suitability score, filtered the result to Italy's boundary and rolled it up, producing a map you can keep exploring. From there you can change the weights, drop land cover or add a dataset and rerun it in seconds.
Nighttime lights in Ukraine, 2016 to 2024. Nine years of annual nighttime lights tell the story through the shape of the curve: it climbs through 2021, then drops by about 65% in 2022. Xun noted that this came from one query against one open dataset, a result Xun said could otherwise take months to assemble.
Summer heat in Europe. The most topical demo took this summer's maximum temperature, joined it to population and counted how many people lived under those temperatures across 10 European countries. That analysis joins a weekly climate grid, a demographic grid and a boundary table, three tables all keyed on the same H3 cell.