GeoSQL is an agent skill designed for data practitioners querying geospatial datasets across PostGIS, BigQuery, Snowflake, and Wherobots. Released under the MIT license and distributed via the dekart-xyz/geosql repository, the tool integrates into Claude, Codex, Claude Code, and GitHub Copilot without requiring a software-as-a-service account.
The skill runs a five-step loop to generate and verify spatial SQL queries. It discovers schema metadata from target tables or public datasets like Overture Maps, generates dialect-specific spatial SQL functions such as ST_INTERSECTS or H3 indexing, and runs geometry validation checks on calculated areas or lengths. For Google BigQuery environments, GeoSQL executes a dry run first to estimate byte volume, enencouraging a default 10 GiB scan ceiling. If a prospective query exceeds this limit, the agent attempts to rewrite it using tighter bounding boxes, lowered H3 resolutions, or extra filters.
Database access relies on local CLI authentication via tools like bq, snow, and dekart, which the README presents as a way to keep warehouse credentials out of the agent context. Visual feedback is enabled optionally through Dekart, an open-source Kepler.gl backend that can run via Docker, self-hosting, or cloud hosting. When Dekart is present, GeoSQL renders the spatial query output, giving the agent an image to inspect for geometry mistakes alongside its textual checks.
Practitioners can install the core package locally using pip install geosql and register it with their chosen assistant using commands like geosql install copilot or Claude Code plugin syntax. Map rendering requires running Dekart and initializing its CLI connection to target data sources.