GeoSQL ships 4x geospatial boost for AI coding agents
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- GeoSQL ships as a skill for Claude, Codex, and GitHub Copilot that targets geospatial workloads on PostGIS, BigQuery, Snowflake, and Wherobots, and runs 100% locally or self-hosted with no SaaS account required.
- GeoSQL claims a 4x improvement on geospatial tasks by putting a rendered map in the agent loop, catching geometry-class errors that text-only validation misses — mistaking a neighborhood polygon for a metro-area perimeter, double-counting overlapping features, or picking the wrong join key on coordinate-reference systems.
- The agent loop enforces BigQuery cost guardrails: every query is dry-run first to estimate bytes scanned, a 10 GiB billing cap is enforced by default, and over-budget queries are rewritten tighter (smaller bbox, lower H3 resolution, more filters) instead of executed.
- Dekart acts as the optional open-source Kepler.gl backend that renders maps for the visual feedback step and can be launched locally with a single
docker run -p 8080:8080 dekartxyz/dekartcommand, with Dekart Cloud as an alternative. - Warehouse credentials never leave the user's machine — GeoSQL piggybacks on local
bq,snow, anddekartCLI authentication rather than passing secrets to the model. - The included eval suite averages 3,085 tokens and 72 seconds per turn, asserts specific behaviors (cost guardrails, validation steps, correct result) rather than just "did the agent answer," and accepts PRs with new test cases.
Why it matters: Data scientists can now plug a Claude/Codex/Copilot agent into PostGIS, BigQuery, Snowflake, or Wherobots without surrendering warehouse credentials, while a hard 10 GiB BigQuery billing cap stops the agent loop from racking up runaway scan costs on an autonomous geospatial workflow.
