GeoSQL Skill Boosts Geospatial Tasks 4x for AI Agents
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- GeoSQL launches as a skill for Claude, Codex, and GitHub Copilot targeting geospatial workflows on PostGIS, BigQuery, Snowflake, and Wherobots, running entirely locally or self-hosted with no SaaS account required.
- The tool claims a 4x improvement on geospatial tasks by inserting a map-rendering step into the agent loop, letting the AI visually catch geometry errors — wrong polygon boundaries, double-counted features, mismatched CRS joins — that text-only validation misses.
- On BigQuery, every query is dry-run first with a 10 GiB billing cap; over-budget queries get rewritten with tighter bounding boxes or lower H3 resolution rather than executed.
- Dekart, the open-source Kepler.gl backend that GeoSQL optionally uses, runs locally via a single
docker runcommand or can be self-hosted on user infrastructure. - Warehouse credentials never reach the agent — GeoSQL relies on local CLI authentication (
bq,snow,dekart), so credentials stay on the user's machine. - The skill ships with a reproducible eval suite under
evals/that asserts cost guardrails, validation steps, and correct results rather than just "did the agent answer"; current benchmarks show an average of 3,085 tokens and 72 seconds per turn.
Why it matters: Data teams already paying for AI coding agents gain a geospatial layer that claims production-grade reliability by catching the geometry-class errors — wrong polygon scopes, double counts, CRS mismatches — that text-only AI agents routinely ship, with warehouse credentials never leaving the user's local CLI. For use cases like EV infrastructure siting, real estate accessibility, and retail site selection, the map-in-loop step is the difference between a demo and a deliverable.
