GIS MCP: real geospatial analysis for agents

Ask a language model how many schools sit within 500 meters of a park and it will probably guess, because it isn’t actually running a buffer and an intersection. GIS MCP connects an agent to the mature Python geospatial stack: the model breaks the question down, and Shapely, PyProj and Rasterio do the geometry, projections and raster math. It had about 200 stars as of 2026-10-06 and is at version 0.15.0, still labeled beta.

What it does

• Vector geometry: 29 Shapely tools such as buffer, intersection, union, voronoi and make_valid, plus length, area and centroids.
• Coordinates and geodesy: transform_coordinates, project_geometry, get_utm_zone, and ellipsoid-aware calculate_geodetic_distance and calculate_geodetic_area.
• Vector tables: GeoPandas file I/O, spatial joins with sjoin_gpd, overlays with overlay_gpd, and attribute dissolves with dissolve_gpd.
• Rasters: clip, resample and reproject, compute_ndvi, and zonal_statistics.
• Spatial stats and data: PySAL autocorrelation and clustering, fetchers for administrative boundaries, climate, ecology, land cover and satellite imagery, and static or interactive web map output.

Who it’s for

• GIS analysts in planning or environmental work who want to run siting and proximity checks in plain language.
• Developers building geospatial agents with LangChain or the OpenAI Agents SDK; the docs include worked examples.
• Researchers who need a quick NDVI or zonal summary without writing a fresh script each time.

Setup

Needs Python 3.10+. Following the README, create a uv virtual environment, then:

uv pip install gis-mcp
gis-mcp

It runs over stdio by default. Point your client at the executable inside the venv:

{
  "mcpServers": {
    "gis-mcp": {
      "command": "/home/YourUsername/.venv/bin/gis-mcp",
      "args": []
    }
  }
}

Install gis-mcp[visualize] for interactive maps. A Docker image is also available and serves HTTP on port 9010.

Our take

Most map servers query some mapping service’s API. GIS MCP is one of the few that computes: it ships no place data and instead lets an agent analyze your own files and layers, which makes it a natural partner to a Google Maps server rather than a rival. The large tool count cuts both ways, since registering everything eats context, so decide which groups you need. HTTP mode adds file upload and download endpoints; put authentication in front before exposing it beyond localhost. MIT licensed.