地理空间MCP服务器

mcp-geospatial

An MCP server for geospatial work: coordinates, geocoding, routing, terrain, weather and radio links. 22 tools, no API key required for any of them.

Built for logistics, telecom, agriculture, insurance, real estate, utilities, emergency response, surveying and mapping — anywhere a question starts with where.

The package and command are mcp-geospatial. The server is not on PyPI under either name: install it from this repository, as below. The PyPI package named geospatial-mcp is an unrelated project; installing it will not give you this server.

"Where can I live within a 30-minute drive of work?"          route_isochrone
"Where's the nearest depot by road to each of today's jobs?"  route_matrix
"How long to drive from the depot to the customer?"           route_directions
"How steep is this section of the trail?"                     terrain_profile
"How far will a 1800 MHz cell reach over suburban ground?"    rf_link_budget
"What's the area of this field in hectares?"                  geom_measure
"Convert this grid reference the survey crew gave me"         coord_convert

Install

Nothing to clone by hand. Every client below runs the server with uvx, straight from this repository. To install it into an environment of your own instead, use pip install git+https://github.com/ystoneman/geospatial-mcp.

claude mcp add geospatial -- uvx --from git+https://github.com/ystoneman/geospatial-mcp mcp-geospatial

{
  "mcpServers": {
    "geospatial": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/ystoneman/geospatial-mcp", "mcp-geospatial"]
    }
  }
}

[mcp_servers.geospatial]
command = "uvx"
args = ["--from", "git+https://github.com/ystoneman/geospatial-mcp", "mcp-geospatial"]

{
  "mcpServers": {
    "geospatial": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/ystoneman/geospatial-mcp", "mcp-geospatial"]
    }
  }
}

{
  "servers": {
    "geospatial": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "git+https://github.com/ystoneman/geospatial-mcp", "mcp-geospatial"]
    }
  }
}

{
  "context_servers": {
    "geospatial": {
      "command": { "path": "uvx", "args": ["--from", "git+https://github.com/ystoneman/geospatial-mcp", "mcp-geospatial"] }
    }
  }
}

docker build -t mcp-geospatial https://github.com/ystoneman/geospatial-mcp.git
docker run -p 8000:8000 mcp-geospatial --transport http

Tools

Every tool taking a location accepts any of these, interchangeably:

48.8584, 2.2945          decimal degrees (canonical, latitude first)
48°51'30"N 2°17'40"E     degrees / minutes / seconds
31UDQ4825211954          USNG / MGRS grid reference
u09tunquc                geohash
geohash:u09tunquc        geohash, marked so it is never taken for a place name
891fb46741bffff          H3 cell
8FW4V75V+9R6             Plus Code
Eiffel Tower, Paris      a place name, geocoded for you

core — 20 tools, on by default

rf — 2 tools, on by default

analysis — 2 tools, opt-in (the stats extra)

● runs with no network at all. ○ calls a free public API.

Link budgets are an everyday tool for mobile operators, wireless ISPs, utilities running SCADA telemetry and IoT fleets: how far a radio reaches for a given power, frequency and environment.

This is one server among many. If you are working out which to use, Sparkgeo's geo-mcp-servers is a curated, machine-checked index of the geospatial MCP field, kept current and marked for liveness — a better place to compare coverage than any claim this README could make about itself. docs/REFERENCES.md lists the servers, standards and research this project leans on or points at.

Selecting toolsets

The examples call the command directly. If it is not installed, prefix it with uvx --from git+https://github.com/ystoneman/geospatial-mcp, as in the client configs above. all also needs the stats extra: uvx --from 'mcp-geospatial[stats] @ git+https://github.com/ystoneman/geospatial-mcp' mcp-geospatial --toolsets all.

mcp-geospatial --toolsets core       # 20 tools, drops the RF set
mcp-geospatial --toolsets all        # adds the analysis toolset (needs [stats])
GEO_TOOLSETS=offline mcp-geospatial  # 13 tools that never touch the network
mcp-geospatial --list-tools          # print the catalogue and exit

offline is for field use with no signal. It keeps only tools that make no network call under any input, so it excludes geom_measure and the place tools — those accept a place name, which means they may geocode.

Configuration

No API key is needed for anything. Every keyed provider has a keyless default.

Data sources and licensing

Results carry the attribution their licence requires in meta.attribution — pass it through when you display them. Two obligations worth knowing before you build on this: OpenStreetMap-derived data (geocoding, places, routing) is ODbL 1.0 and requires attribution, and Open-Meteo's free tier is non-commercial, so commercial deployments need OPEN_METEO_API_KEY or a self-hosted instance.

The public instances are free and shared. This server caches aggressively and rate-limits per host to stay inside their usage policies; for heavy use, self-host and set the corresponding *_URL. See Sources and credits for the complete list.

What it gets right that is easy to get wrong

• Distances and areas are geodesic. Karney's algorithm on the WGS84 ellipsoid, accurate to millimetres at any distance. Not 111.32 km/degree, which is 0.25% out at 60° latitude and worse further north.
• Buffers are circles on the ground. Buffering in degrees produces an ellipse stretched by 1/cos(latitude) — 2× wrong at 60°, 5× at 78°. Every buffer here is projected to an azimuthal equidistant plane first.
• Grid references truncate, they never round. A reference names the square a position falls in, so dropping digits must move the point to the coarser square's corner. Rounding would hand back a neighbouring square — a kilometre away at four digits, and confidently wrong.
• DEM gaps stay gaps. A missing elevation is reported as null, never substituted with zero. Substituting sea level flattens a ridge into an entirely plausible plain, and every statistic downstream inherits it.
• Place names are never silently guessed. "Springfield" matches cities of near-equal prominence 1,500 km apart, so a tool asked about it lists them and asks, rather than quietly answering for Illinois. When a name is clear enough to use, meta.notes says exactly which place it was taken to mean, and names any others that share it.
• A place name is not mistaken for a code. "Bergen", "Denver" and "Sydney" are spelt entirely from the geohash alphabet, and decoding them lands in Alaska, the Atlantic and Turkey. A bare token is decoded as a geohash at once only when it is written like one, in lowercase with a digit among the letters. Anything else that could be a geohash is geocoded first, with a note in meta.notes giving the geohash reading, and is read as a geohash only if no place has the name. Write geohash:gcpvj to skip the lookup. A 4-character geohash such as sw19 could be a postcode district, so it is decoded with a note saying so.
• Errors are errors. Failures raise, so the MCP response is marked isError. A tool that returns {"status": "error"} is reported to the model as a success, and it will proceed on invented data.
• Empirical models report their limits. COST231-Hata outside 1500–2000 MHz still returns a number; meta.notes tells you it is out of range.
• CRS transforms admit their accuracy. Every transform names the operation PROJ used for that point and its published accuracy. When a better operation needs a grid file that is not installed, the response names the file; when no published transformation exists at all, it says the datum shift was ignored.

Development

git clone https://github.com/ystoneman/geospatial-mcp && cd geospatial-mcp
uv sync --all-extras
make check          # lint, types, tests — exactly what CI runs
make evals          # tool-selection eval suite
make evals-network  # the same, plus the cases that call live APIs

See AGENTS.md for the layout and conventions, and CONTRIBUTING.md to add a tool.

Sources and credits

Almost none of the geospatial computation here is ours. This package is an integration layer: it chooses which established libraries, published algorithms and public data services to expose to a language model, wires them together, and tries hard to get the units, datums and error handling right. The interesting work was done by the projects below, and this section exists so you can see exactly which ones you are depending on.

To put a number on it — of the Python that ships when you install this package, 1.0% is ours and 99.0% is dependencies (6,912 lines against 654,791), inside a 150 MB installation of which 109 MB is compiled native libraries and 9 MB is the EPSG registry. Counted by installing the package with no extras into an empty environment and measuring site-packages.

Runtime dependencies

Everything below is permissively licensed. Licences were read from each package's own LICENSE file, not from its classifiers — the two disagree more often than you would expect.

Optional extras: pillow (MIT-CMU), sgp4 and skyfield (MIT), scikit-learn, esda and libpysal (BSD-3-Clause), fast-tsp (MIT), staticmap (Apache-2.0), rasterio and rio-tiler (BSD), pystac-client (Apache-2.0), morecantile (MIT).

Native libraries bundled inside those wheels

GEOS is the one copyleft component in the tree. It is unavoidable — it sits under Shapely, GeoPandas, OGR, PostGIS and DuckDB-spatial alike, and there is no permissive replacement with equivalent overlay correctness.

Live data services

None require an API key. Results carry the attribution their licence demands in meta.attribution; pass it through when you display them.

Algorithms, and where they come from

Every computation in this package traces to a library, a published formula or a public service. None of it was invented here, and the literature is cited so you can check the implementation against the source.

"Implemented by: this package" means the formula was written out from the cited publication in geo/rf.py or geo/terrain.py, because no permissively licensed Python library implements it. The mathematics is not ours; only the transcription is.

What this package actually adds

Given the above, it is worth being plain about what is left:

• The composition — which capabilities belong in one server, and what the tool surface should look like for a model rather than a human.
• Correct units and datums at the seams. Geodesic buffering via an azimuthal equidistant reprojection, geodesic path sampling, geodesic area — the places where naive degree arithmetic silently gives wrong answers.
• A shared HTTP layer with per-host rate limiting, a disk cache, bounded retries and mirror failover, so the free public services are used within their stated policies.
• Error design — tools raise rather than returning an error payload, and every message names the input, the problem and a valid example.
• Attribution plumbing, so licence obligations travel with the data.
• Tests and evals — 293 tests and a 31-case tool-selection eval library.

Full licence texts and obligations are in THIRD_PARTY_NOTICES.md. If you spot a missing or incorrect attribution, please open an issue — that is a bug.

Licence

MIT — see LICENSE.