Summary

GeoSQL transforms LLMs (Claude, Codex, Copilot) into geospatial analytics agents for data scientists and analysts. It enables efficient querying of geospatial data across PostGIS, BigQuery, Snowflake, and Wherobots, offering significant performance improvements locally or self-hosted, without SaaS.

How It Works

GeoSQL uses an agent loop with map-in-the-loop feedback. It discovers warehouse metadata, generates engine-specific spatial SQL, and performs cost checks (e.g., BigQuery dry-runs with a 10 GiB cap, query rewriting). Geometry validation and map rendering via Dekart allow the agent to self-correct geometric errors missed by text-only analysis, achieving a reported 4x improvement. Warehouse credentials remain local via CLI authentication.

Quick Start & Requirements

Install with pip install geosql. Install agents: geosql install claude, geosql install codex, geosql install copilot. For local Dekart: docker run -p 8080:8080 dekartxyz/dekart, then pip install dekart && dekart init to connect databases (PostGIS, BigQuery, Snowflake, Wherobots). Requires Python, Docker (for local Dekart), and CLI auth for warehouses.

Highlighted Details

• Achieves 4x performance improvement via map-in-the-loop visual feedback.
• Supports PostGIS, BigQuery, Snowflake, and Wherobots.
• Includes a reproducible evaluation suite (evals/) for rigorous testing.
• Operates 100% locally or self-hosted, no SaaS required.

Maintenance & Community

No specific contributors, sponsorships, or community channels (Discord/Slack) are detailed. Contributions to the evaluation suite are welcomed via PRs.

Licensing & Compatibility

The license type and compatibility notes for commercial use are not specified in the provided README text.

Limitations & Caveats

Performance gains rely on map-in-the-loop; text-only interactions may be less accurate. Local Dekart installation is recommended for full map visualization. The project's license is not explicitly stated.