Query geospatial databases and render validation maps

GeoSQL is a Claude, Codex, and GitHub Copilot skill for data scientists and analysts running geospatial queries on PostGIS, BigQuery, Snowflake, and Wherobots, entirely locally or self-hosted with no SaaS account required. It runs a five-step agent loop with a map inside it: discovery, exploring the warehouse's own metadata - tables, columns, types - instead of guessing schemas, working with Overture Maps shares on BigQuery and Snowflake plus private tables on any of the four supported engines; SQL generation, writing spatial SQL using the correct functions for the target engine such as ST_INTERSECTS, ST_DISTANCE, H3, and bbox overlap for partition pruning; a cost check on BigQuery specifically, where every query is dry-run first to estimate bytes scanned against a default 10 GiB billing cap, and an over-budget query gets rewritten cheaper (tighter bbox, lower H3 resolution, more filters) rather than executed; geometry validation, computing total area for polygons or total length for lines as a sanity check against domain knowledge; and map feedback, where the agent renders the result through Dekart when available, looks at the rendered image, and corrects geometry mistakes a text-only loop would miss - specifically the step credited with a measured 4x improvement on geospatial tasks. Warehouse credentials never reach the agent, since the skill authenticates through the user's own local CLI tools (bq, snow, dekart).

Use it inside Claude, Codex, or GitHub Copilot for geospatial analysis prompts against PostGIS, BigQuery, Snowflake, or Wherobots data - real estate accessibility analysis, site selection by point-of-interest co-location, or infrastructure density mapping, with documented example prompts such as "Show buildings with low school accessibility in Ottawa, render as a map" and "Find the top 10 locations for Sporting Goods Store in Seattle based on POI co-location and distance to the nearest competitor." Map rendering, reading maps back to fix geometry, and connecting to the four supported warehouses all specifically require the optional Dekart component; without Dekart the skill still works but loses the map-feedback step that produces the 4x improvement. Dekart itself connects either to Dekart Cloud, which needs no Docker, or to a local or self-hosted instance.

Input is a natural-language geospatial request via the /geosql prompt; output is spatial SQL executed against the target warehouse, a geometry-validated result, and, when Dekart is configured, a rendered map. GeoSQL ships a reproducible eval suite under evals/ asserting specific behaviors - cost guardrails, validation steps, correct results - not just whether the agent produced an answer; the included suite, covering london-boroughs, berlin-create-map, and paris-boundaries with 8 assertions total, currently passes 100%, averaging 3,085 tokens and 72 seconds per turn. The suite is runnable directly with python evals/run.py, with setup and instructions for adding new cases documented in evals/RUNBOOK.md.

Installs via pip install geosql && geosql, which detects the agent, installs the skill, and only then checks whether the optional Dekart CLI is ready, offering to run python -m pip install dekart and dekart init if not, or printing those exact commands for later if declined. Supported data warehouses are PostGIS, BigQuery, Snowflake, and Wherobots, authenticated through each engine's own local CLI (bq, snow) rather than credentials passed to the agent; map rendering and the read-back validation loop go through the separate Dekart project, which can be self-hosted or run through Dekart Cloud.

Data scientists and analysts who want an AI agent to write correct spatial SQL and catch its own geometry mistakes, via an actual rendered map rather than just a text sanity check, against PostGIS, BigQuery, Snowflake, or Wherobots, without sending warehouse credentials to the agent or a third-party SaaS account.