geo-explorer-mcp

An MCP server that serves country facts, administrative boundaries and regional statistics as structured data, so an AI assistant can build interactive geography instead of reciting it.

The server deliberately returns data, not prose. It hands back populations, polygons and coordinates; the model decides how to present them — as a clickable map, a comparison table, or a lesson in Slovak, German or Hungarian. Nothing in this server knows how to translate, and it does not need to.

Tools

country accepts a name in almost any language — Slovakia, Slovensko, Magyarország, Magyarorszag (no diacritics), Deutschland, Ungarn.

The two boundary and statistics tools are designed to be joined: get_region_details returns a match_key per region, and a join_hint telling the model how to normalise get_map_data's names to match it.

Quick start

git clone https://github.com/arnienemeth/geo-explorer-mcp
cd geo-explorer-mcp
uv sync

Copy .env.example to .env and fill it in:

RESTCOUNTRIES_API_KEY=rc_live_...          # free: https://restcountries.com/sign-up
GEO_EXPLORER_CONTACT=https://github.com/you/your-repo

GEO_EXPLORER_CONTACT goes into the User-Agent sent to Wikidata. Wikimedia requires a contact URL or email and answers anything else with 403.

Once published to PyPI it also installs with no clone at all:

uvx geo-explorer-mcp

Try it in the MCP Inspector:

uv run fastmcp dev inspector server.py

Claude Desktop

Add this to claude_desktop_config.json (on Windows, %APPDATA%\Claude\claude_desktop_config.json), adjusting the paths:

{
  "mcpServers": {
    "geo-explorer": {
      "command": "C:\\path\\to\\geo-explorer-mcp\\.venv\\Scripts\\python.exe",
      "args": ["C:\\path\\to\\geo-explorer-mcp\\server.py"]
    }
  }
}

Pointing at the project's own .venv interpreter means Claude runs exactly what you tested, with no second environment to keep in sync. Restart Claude Desktop fully afterwards — the config is only read at startup.

Then ask it something like:

Show me an interactive map of Slovakia's regions, shaded by population, in German.

Demo page

demo/index.html renders any of four countries from the same two sources the server uses, joined in the browser: shading by population, area or density, capital markers, a sortable table, and per-region links to Wikipedia and Wikidata. One page, parameterised by country — because the server is generic.

It must be served over http, not opened as a file — a file:// page has a null origin and the browser blocks its cross-origin fetches:

cd demo
python -m http.server 8000

The single US miss is the Virgin Islands, which has no ISO 3166-2 entry in Wikidata. Regions without a match render grey rather than being dropped.

Layout

server.py                      entry point kept at the root (a shim)
src/geo_explorer_mcp/server.py the implementation
demo/index.html                the browser demo
server.json                    MCP registry metadata

server.py at the root re-exports the server object, so the Claude Desktop config, fastmcp dev inspector server.py and the probe scripts all keep working while the package underneath stays publishable.

Tests

uv run python probe2.py        # all three tools, plus the boundary/statistics join
uv run python probe_names.py   # country-name resolution across languages and spellings

probe2.py runs all four demo countries and asserts two things that previously broke: that geojson_url is the resolved Git LFS media URL, and that at least 90% of boundary regions join to their statistics. Germany sat at 50% before native-name matching landed.

Field notes

Things that cost real debugging time, written down so they don't cost yours.

REST Countries v3.1 is gone. Nearly every tutorial online still uses it. v5 lives at a different host, requires a bearer token, and renamed every field (area → area.kilometers, cca3 → codes.alpha_3, and capitals is now an array of objects, not strings).

geoBoundaries GeoJSON is stored in Git LFS, and only media.githubusercontent.com serves it to a browser:

• raw.githubusercontent.com/... returns a 131-byte LFS pointer file, not geometry.
• The github.com/.../raw/... URL the API publishes 302-redirects with an empty access-control-allow-origin. Browsers enforce CORS on every hop of a redirect, so a page fetching it fails with TypeError: Failed to fetch even though the final response sends access-control-allow-origin: *.

Either mistake produces a blank map with no error. get_map_data resolves the redirect and returns the URL that actually works.

Verifying with curl does not verify browser behaviour. curl -L followed that redirect happily and reported the permissive header on the final response. It ignores CORS entirely. Test cross-origin fetches with an Origin header, or in a browser.

Wikidata's area property mixes units. P2046 values are entered in square kilometres, hectares or square metres, and the raw number carries no hint which. Reading it directly gave Békéscsaba an area of 193,930,000 km² — it is 193.9 km², recorded in m². Use the normalised value (p:P2046/psn:P2046), which Wikidata converts to SI base units.

ISO 3166-2 mixes administrative levels. Hungary returns 43 subdivisions: 19 counties, Budapest, and 23 cities with county rights. The cities sit inside the counties, so their areas and populations must not be summed, and there are far fewer boundary shapes than entries. get_region_details returns subdivision_types and warns via data_quality_notes.

Country-name matching folds case but not diacritics. Magyarország resolves; Magyarorszag does not. Translations (Ungarn, Hongrie) are excluded from the name aggregate on purpose. The server works around both with a three-step fallback: the name aggregate, then the translations endpoint, then a locally built diacritic-folded index.

A Wikidata query that works for a small country can time out for a large one. Resolving subdivisions via ?i wdt:P17 ?country makes Wikidata scan everything in that country: fine for Slovakia, HTTP 504 for the United States. Filtering on the ISO 3166-2 prefix instead — FILTER(STRSTARTS(?iso, "US-")) — returns the same 56 rows in 0.9 seconds, because those codes are defined as <alpha-2>-<subdivision>, so the prefix already is the country filter.

The same region has different names in each source. geoBoundaries says Bayern, Sachsen, Thüringen; Wikidata's English labels are Bavaria, Saxony, Thuringia. Matching on the English label alone joined 8 of Germany's 16 states. Building a key from every name variant, native names included, joins all 16 — which is why get_region_details returns match_keys (plural).

A bounding box lies about shapes that cross the antimeridian. Alaska's Aleutian Islands run past 180°, so its box reads -179.15 .. 179.78 — a 359-degree span. fitBounds on that zooms out to the whole globe and shrinks the mainland to a smudge, with no error. Russia, Fiji and New Zealand share the trap. The demo refuses any box wider than 180° and falls back to a fixed view.

The hidden attribute loses to any author display rule. It works through the browser's default stylesheet, the weakest source there is, so an element styled display:grid stays visible when you set hidden. Symptom: a loading overlay that never goes away, covering a map that loaded perfectly. Guard it with [hidden]{display:none !important}.

Some upstream region names are damaged. geoBoundaries' ADM2 names for Slovakia are truncated and mistransliterated (Prešov → Predov, Dolný Kubín → Dolne Kub). The geometry is fine. get_map_data detects this and returns a data_quality_note so a model does not present corrupted names as fact.

No # comments inside a SPARQL query that gets collapsed to one line. In SPARQL # comments out the rest of the line; after collapsing, that is the entire query. Symptom: HTTP 400.

Data sources