A Model Context Protocol (MCP) server lets an AI agent geocode an address in a second. Ask the same agent for the next bus leaving the stop outside, and the catalogue we maintain has no server that can answer.
That gap is typical of where geospatial MCP servers stand today. Some areas have a dozen entries that overlap. Others have none.
Coverage matters because, in geography, 90% right is 100% wrong, as I wrote on the Sparkgeo blog. An agent that names a nearby hospital instead of the nearest one has failed, however close it lands. A server that answers from real data gives the agent a control point it cannot argue with.
The geo-mcp-servers repository tracks these servers and health checks them. I wrote an analysis of the catalog as it stood on October 15, 2026. This post covers the parts that matter if you choose servers for your agents, and the parts that matter more if you build them.
Where the servers are
Everyday location tasks have more servers than anyone needs. 21 servers carry the geocoding tag. Seven of the 17 in the geocoding category wrap OpenStreetMap services such as Nominatim or Overpass, and five of the 17 are stale. I would not build another OpenStreetMap geocoder. A new one earns its place with different data or a different way of working: Overture places, US address verification, or lookups that run offline from a local file.
The large mapping platforms have arrived, with gaps. 12 of the 18 platform servers are official, including Mapbox, TomTom, MapQuest, CARTO, Felt, Cesium and an Esri beta. Google Maps has four entries and none is official, and the reference server is archived. Esri has seven entries across categories, and one is an official beta. Official status does not guarantee upkeep either: three official platform servers are stale.
Analysis tooling is the thin part. Spatial databases (3 servers), desktop GIS (9) and geoprocessing toolkits (7) together make up less than a fifth of the catalogue. No desktop GIS server is official. All five QGIS servers and both ArcGIS Pro servers come from the community. An agent that does GIS work beyond lookups has the fewest options here.
Weather is open and heavily US-based. All 12 weather servers are open or free, and most draw on NOAA, the US National Weather Service or Open-Meteo. Canada is the only other country with a national service in the catalog. ECMWF open data, DWD and the Met Office do not appear, apart from ERA5 data through Open-Meteo.
Where servers run, and why it matters for trust
Deployment decides who can use a server and who sees your queries. Only about a third of the servers work from a web client such as claude.ai without an install.
| Deployment | Servers | Official | Community | What it means for you |
|---|---|---|---|---|
| Local | 53 | 12 | 41 | Install it and your client starts it on your machine. Needs a desktop client. |
| Hosted | 24 | 22 | 2 | Connect to the vendor’s URL. Works from web clients, and queries go to the vendor. |
| Local + hosted | 15 | 5 | 10 | Run it yourself, or use a public endpoint someone else runs. |
| Self-hosted | 5 | 0 | 5 | Deploy it as an HTTP service yourself. |
| Add-in | 1 | 0 | 1 | Runs inside a desktop application (ArcGIS Pro). |
| Unknown | 8 | 2 | 6 | The README does not say. |
Vendors host and the community ships local servers. 22 of the 41 official servers are hosted, and 41 of the 65 community servers run locally. I expect hosted to become the default delivery model for commercial geodata, because it keeps the vendor’s data and API keys on the vendor’s side. It also removes one check we can make. A hosted server has no source code to scan, so trust rests on the vendor’s domain.
Public endpoints need the same care. Two community operators run most of them: Pipeworx for six servers and cyanheads for four. These are convenient. Your queries then pass through a third party that is neither you nor the data provider. Check who runs the domain before you connect.
Nine gaps with no dedicated server
Nine capabilities have no dedicated server in the catalogue. Some may exist and not be listed yet, so search before you treat any of these as confirmed.
| Capability | What the catalog has today |
|---|---|
| Public transit (GTFS, schedules, real-time) | Nothing |
| Fleet and logistics optimisation (vehicle routing, dispatch) | Trip optimisation inside OSRM MCP only |
| Cloud warehouse spatial SQL (BigQuery, Snowflake, Databricks) | Wherobots (Apache Sedona), plus warehouse access through CARTO and Felt |
| Natural hazards (wildfire, flood, earthquake) as a focus | Wildfire inside weather-mcp; earthquakes inside God’s Eye View and the EarthScope server |
| Demographics and census data | Demographics derived from mobility data in Factori (paid), and census enrichment inside Smarty |
| Elevation and terrain as a service | Spread across MapTiler, TerraLab, MapSmith and Open-Meteo |
| 3D, point clouds and digital twins | Cesium and one module of the Kiro power pack |
| Parcels and property outside the US and Canada | Nothing |
| National weather services outside North America | Nothing (see above) |
Transit, non-North American parcels and non-North American weather services have nothing at all. The other six have partial coverage, spread across servers that were built for something else.
Where to build next, and how to use the catalogue
If you build servers, pick from the gap table or from analysis tooling. Desktop GIS automation, cloud warehouse spatial SQL, transit and hazard data have the clearest demand and the least supply. Another OpenStreetMap geocoder has neither.
If you choose servers, treat commodity lookups as settled and choose on data quality, licence and vendor support. Prefer official or actively maintained servers. Check the deployment value first: in a web client, only hosted and local + hosted servers will work.
The repository also includes a skill for finding the server you need. It lives in the repository’s skill folder. Ask it whether an MCP server exists for a task such as geocoding, PostGIS, STAC imagery or weather, and it searches the catalog and recommends options. When you find a server the catalog is missing, the same skill adds it to the tracked list.
Sparkgeo builds on this protocol too. Prescient, our cloud-native geospatial infrastructure platform, has an MCP server of its own.
The tables in the repository README are generated and stay current. The analysis is a snapshot from October 15, 2026, so re-run the numbers before you rely on them later. If you build a server for one of the nine gaps, or know one we have missed, open an issue or a pull request on geo-mcp-servers.
