Geo SQL is a new tool designed to transform large language models like Claude and Codex into specialized geospatial analytics agents. It integrates with popular data warehouses including Post GIS, Big Query, and Snowflake, as well as Wherobots, enabling data scientists and analysts to query geospatial data using natural language. A significant advancement is its "map-in-the-loop" capability, which reportedly yields a 4x improvement in geospatial task accuracy.

This tool operates entirely locally or can be self-hosted, eliminating the need for SaaS accounts. Users can install Geo SQL directly into agents or via plugins. It leverages Dekart, an open-source backend, for map rendering and database connections. Geo SQL intelligently writes spatial SQL queries, incorporating engine-specific functions and performing cost checks, especially on Big Query where it estimates query costs and rewrites queries to stay within a 10 GiB billing cap.

The "map-in-the-loop" mechanism allows the AI agent to visualize query results, identify geometry-related errors that text-only validation misses, and self-correct. This closed-loop system, incorporating visual feedback, is the core of its performance gains. Geo SQL includes a reproducible evaluation suite to verify specific behaviors like cost guardrails and result correctness, with current benchmarks showing 100% pass rates on tested assertions.