About Bbox Mcp Server
Bbox Mcp Server provides geospatial tools for querying OpenStreetMap data, converting bounding boxes, and analyzing spatial distributions. It serves AI agents and GIS users with structured results and map links for visual verification.
Key features
- 01Convert bounding boxes across formats and projections
- 02Generate H3 cells for a bounding box
- 03Search OpenStreetMap through Overpass
- 04Look up curated OpenStreetMap tags
- 05Aggregate query results into H3 cells
- 06Create shareable map links
Use cases
- 01Find nearby points of interest on OpenStreetMap
- 02Convert WKT or GeoJSON bounding boxes to another projection
- 03Compare spatial density across an area
- 04Check query results on an interactive map
MCP support
It exposes six geospatial tools to MCP clients, including bounding-box conversion, Overpass queries, H3 aggregation, and map-link generation.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y bbox-mcp-serverMAPBOX_ACCESS_TOKENSecretMapbox token for natural-language location search (e.g. 'San Francisco')OVERPASS_API_URLCustom Overpass API endpoint URL (defaults rotate between public instances)MAX_H3_CELLSSafety cap for H3 grid generation (default: 50000)Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
7Saved tool list · updated 3w ago
| Parameter | Type | Required | Description |
|---|---|---|---|
| location | string | No | A text location to search for (e.g. 'New York City'). Requires MAPBOX_ACCESS_TOKEN env var. Either 'location' or 'bbox' MUST be provided. |
| bbox | string | No | The geometry to parse. Can be a raw bounding box string ('lat1,lng1,lat2,lng2'), a WKT polygon/linestring/point, a GeoJSON payload, or an `ogrinfo` extent block. The Minimum Bounding Rectangle (MBR) encapsulating the geometry will be extracted. Either 'location' or 'bbox' MUST be provided. |
| epsg | string | No | The projected EPSG code (e.g. '3857'). Defaults to '4326' (WGS84). Over 3,900 bundled projections; unknown codes are auto-fetched from epsg.io. |
| format | enum | No | Format of the output. Options: wkt, geojson-bbox, leaflet, overpass, ogc-bbox, kml, geojson-polygon, csv, stac-bbox. Default: csv.wkt | geojson-bbox | leaflet | overpass | ogc-bbox | kml | geojson-polygon | csv | stac-bbox |
| coord_order | enum | No | Toggle between 'lng,lat' (default) and 'lat,lng' coordinate ordering in the formatted output. Useful for APIs that expect swapped coordinate orders.lng,lat | lat,lng |
| zoom | number | No | The map zoom level (0-22) used to calculate the map tile coordinates for the centroid. Defaults to 15. |
| precision | number | No | The number of decimal places to output coordinate strings as (defaults to 6). |
Details
From the project README🌍 bbox-mcp-server
Ask AI about anything, anywhere — and verify the answer on a map. For Free
The geospatial toolkit for AI agents. 6 tools, zero config — give any LLM the ability to find, query, convert, and aggregate spatial data using open data. No API keys or signups required to start.
Every response includes a shareable verification link to vibhorsingh.com/boundingbox — click it to visually confirm results on an interactive map. No other MCP server does this.
What Can You Do With This?
You don't have to be a GIS professional to make use of this. Any AI agent with this MCP server can answer spatial questions using real OpenStreetMap data and not spit out hallucinated garbage.
| Ask your AI agent... | What happens under the hood |
|---|---|
| "How many EV chargers are in downtown Denver?" | Overpass query → H3 hex binning → density analysis |
| "Is there a hospital near this Airbnb?" | POI search with structured tags → map verification link |
| "Find all playgrounds within 1km of this address"* | Overpass query + radius filter → pinned results on a shareable map |
| "Convert this WKT to GeoJSON in EPSG:3857" | Format conversion across 6 inputs, 9 outputs, 3,900+ projections |
| "Show me all bike-share stations in Amsterdam" | Curated OSM tags → Overpass query → results on map |
| "Compare park density across Seattle neighborhoods" | Overpass + H3 aggregation → hex-binned spatial analysis |
Every answer comes with a link. Click it, see if the AI got it right and you can also share it with someone else.
What Makes This Different
Most AI tools give you text you have to trust. This one gives you an interactive map you can check and verify!
- Verifiable — Every response includes a public URL where you (or anyone) can visually confirm the results. This is the only MCP server that does this in any domain.
- Deterministic — Queries hit real OpenStreetMap data and return real coordinates. The AI interprets your question. The data comes from all the hard work done by all the awesome OSM volunteers.
- Free — No paid services. Overpass API is free, OSM data is free, the tool is free. A Mapbox token (free tier, no payment info) unlocks natural language location search but isn't required.
- Zero config —
npx -y bbox-mcp-serverand you're running.
Why This Exists
| The problem | How bbox-mcp solves it |
|---|---|
| "I have WKT but the API needs a GeoJSON bbox in EPSG:3857." | Parses 6 input formats, projects to 3,900+ EPSG codes, outputs in 9 formats — in one call. |
| "I keep getting the wrong OSM tags for Overpass queries." | list_osm_tags returns curated tag combos. No more hallucinated amenity=grocery. |
| "How many hospitals are in this district?" | aggregate_overpass_h3 queries Overpass and bins results into H3 hexagons server-side. |
| "Is this bounding box actually correct?" | Every response includes a clickable map link for visual verification. |
Tools at a Glance
| Tool | What it does | Key params |
|---|---|---|
get_bounds | Convert and project a bbox across formats and coordinate systems | bbox, epsg, format, coord_order, zoom |
get_h3_indices | Generate H3 hex cell indices covering a bbox | bbox, resolution, compact |
generate_share_url | Create a shareable map link for a bbox | bbox |
search_overpass | Query OpenStreetMap via Overpass QL within a bbox or radius | bbox, query, limit, radius_meters |
list_osm_tags | Look up correct OSM tags for a category | category |
aggregate_overpass_h3 | Run an Overpass query and bin results into H3 hexagons | bbox, query, resolution |
All tools accept location (natural language, requires Mapbox token) or bbox (coordinates, WKT, GeoJSON, etc).
Quick Start
Add to your MCP client config:
Claude Desktop (Click to expand)
Add to claude_desktop_config.json:
{
"mcpServers": {
"bbox": {
"command": "npx",
"args": ["-y", "bbox-mcp-server"]
}
}
}
Cursor (Click to expand)
Add to .cursor/mcp.json:
{
"mcpServers": {
"bbox": {
"command": "npx",
"args": ["-y", "bbox-mcp-server"]
}
}
}
Windsurf (Click to expand)
Add to ~/.codeium/windsurf/mcp_config.json:
{
"mcpServers": {
"bbox": {
"command": "npx",
"args": ["-y", "bbox-mcp-server"]
}
}
}
VS Code (GitHub Copilot) or Google Antigravity (Click to expand)
Add to .vscode/mcp.json in your workspace:
{
"servers": {
"bbox": {
"command": "npx",
"args": ["-y", "bbox-mcp-server"]
}
}
}
Try: "Find all coffee shops within 500m of Times Square"
Optional Configuration
{
"mcpServers": {
"bbox": {
"command": "npx",
"args": ["-y", "bbox-mcp-server"],
"env": {
"MAPBOX_ACCESS_TOKEN": "pk.your-token-here",
"OVERPASS_API_URL": "https://your.custom.overpass.instance/api/interpreter",
"NOMINATIM_API_URL": "https://your.custom.nominatim.instance"
}
}
}
}
| Variable | Default | Description |
|---|---|---|
MAPBOX_ACCESS_TOKEN | — | Enables natural language location search (e.g. "San Francisco") |
OVERPASS_API_URL | auto | Custom Overpass endpoint. By default, rotates between overpass-api.de and kumi.systems. |
NOMINATIM_API_URL | https://nominatim.openstreetmap.org | Custom Nominatim endpoint. Set this if hosting as a shared/remote MCP server or using a commercial provider. |
MAX_H3_CELLS | 50000 | Safety cap for H3 grid generation |
Or install globally: npm install -g bbox-mcp-server
Tool Reference
get_bounds
Convert and project a bounding box across 6 input formats, 9 output formats, and 3,900+ coordinate systems. Returns the center point and tile coordinates for the centroid.
| Param | Type | Default | Description |
|---|---|---|---|
bbox | string | — | Input geometry (coordinates, WKT, GeoJSON, ogrinfo extent) |
epsg | string | "4326" | Target projection. Unknown codes auto-fetched from epsg.io. |
format | string | "csv" | Output: csv, wkt, geojson-bbox, geojson-polygon, leaflet, overpass, ogc-bbox, kml, stac-bbox |
coord_order | string | "lng,lat" | Swap to "lat,lng" for APIs that expect latitude first |
zoom | number | 15 | Zoom level for tile coordinate calculation |
precision | number | 6 | Decimal places in formatted output |
💡 Prompt: "Get the bounding box for Central Park in WKT format projected to EPSG:32618"
get_h3_indices
Generate Uber H3 hexagonal cell indices covering a bounding box.
| Param | Type | Default | Description |
|---|---|---|---|
bbox | string | — | Input geometry |
resolution | number | — | H3 resolution (0–15) |
compact | boolean | false | Merge cells into coarser parents where possible |
return_geometry | boolean | false | Include GeoJSON hex boundaries |
💡 Prompt: "Give me H3 cells at resolution 7 for downtown Chicago, include the hex geometries"
search_overpass
Execute an Overpass query within a bounding box or radius. Returns structured results with names, coordinates, and tags. The verification link plots each result as a pin on the map.
| Param | Type | Default | Description |
|---|---|---|---|
bbox | string | — | Input geometry |
query | string | — | Overpass QL (e.g. nwr["amenity"="cafe"]). The server wraps it in the appropriate spatial filter automatically. |
radius_meters | number | — | Search radius in metres. Use only when the user specifies an explicit distance (e.g. 2000 for "within 2km"). Omit for area/region queries. |
limit | number | 100 | Max elements returned |
💡 Prompt: "Find all parking within 2km of JFK airport"
💡 Prompt: "Search for hospitals in Manhattan, limit 50"
list_osm_tags
Look up the correct OpenStreetMap tags for a category before writing an Overpass query.
| Param | Type | Description |
|---|---|---|
category | string | Broad category (e.g. "food", "health", "transport") |
💡 Prompt: "What are the correct OSM tags for supermarkets?"
aggregate_overpass_h3
Run an Overpass query and bin results into H3 hexagons for spatial density analysis. Returns counts per cell and GeoJSON hex boundaries.
| Param | Type | Default | Description |
|---|---|---|---|
bbox | string | — | Input geometry |
query | string | — | Overpass QL core query |
resolution | number | 8 | H3 resolution for binning |
💡 Prompt: "Aggregate all hospitals in Seattle into H3 bins at resolution 7"
generate_share_url
Generate a shareable link to visualize a bounding box on the interactive map at vibhorsingh.com/boundingbox.
| Param | Type | Description |
|---|---|---|
bbox | string | Input geometry |
💡 Prompt: "Generate a share link for bbox 40.7128,-74.0060,40.7580,-73.9855"
Supported Input Formats
All tools auto-detect the input format. No need to specify which one you're using.
| Format | Example |
|---|---|
| Raw coordinates | 40.7128,-74.0060,40.7580,-73.9855 |
| WKT | POLYGON((-74.006 40.712, -73.985 40.712, ...)) |
| GeoJSON | {"type":"Feature","geometry":{...}} |
| GeoJSON bbox | {"bbox":[-74.006,40.712,-73.985,40.758]} |
| ogrinfo extent | Extent: (-74.006, 40.712) - (-73.985, 40.758) |
| Space-separated | 40.7128 -74.0060 40.7580 -73.9855 |
🤖 For AI Agent Developers
Response structure
Every tool returns two content blocks:
- Human-readable text — formatted output with the map verification link
- Structured JSON — all computed data, machine-parseable
Example get_bounds JSON response:
{
"original_wgs84": { "lat1": 40.7128, "lng1": -74.006, "lat2": 40.758, "lng2": -73.9855 },
"projected": { "xmin": -8238310.23, "ymin": 4970241.32, "xmax": -8235527.11, "ymax": 4976491.56 },
"center": { "lat": 40.7354, "lng": -73.99575 },
"tile_indices": { "z": 15, "x": 9660, "y": 12284 },
"epsg": "3857",
"coord_order": "lng,lat",
"area_km2": 8.681,
"dimensions": { "width_km": 1.714, "height_km": 5.066 },
"share_url": "https://vibhorsingh.com/boundingbox/#40.712800,-74.006000,40.758000,-73.985500"
}
Error handling
All errors return isError: true with a descriptive message. Invalid coordinates, unknown EPSG codes, and oversized H3 requests all return clean errors — the server never crashes on bad input.
Logging
Structured JSON logs go to stderr (stdout is reserved for MCP protocol). Each entry includes timestamp, level, and context.
Acknowledgments & Fair Use
OpenStreetMap
This tool is built on OpenStreetMap: A global dataset created and maintained by millions of volunteers. Every query you run returns data that someone walked, mapped, or verified by hand. If you find this useful, consider contributing to OSM.
Responsible Usage
The public Overpass and Nominatim API instances are free community resources with limited capacity. Avoid tight loops, excessive polling, or bulk-scraping. If hosting this as a shared/remote MCP server, point NOMINATIM_API_URL and OVERPASS_API_URL at your own instance or a commercial provider (see Overpass API installation and Nominatim installation). The default server rotation for Overpass helps spread load, but it's not a substitute for responsible usage.
License
MIT
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.