myosm-mcp-server — OpenStreetMap MCP Server (Node.js)

An OpenStreetMap MCP server that enhances LLM capabilities with location-based services and geospatial data.

This is a Node.js/TypeScript reimplementation of jagan-shanmugam/open-streetmap-mcp (Python), extended with map-layer, public-transport and energy-infrastructure tools.

Features

The server gives LLMs tools to interact with OpenStreetMap data:

• Geocode addresses and place names to coordinates (and the reverse)

• Find nearby points of interest

• Get route directions and commute analysis between locations (OSRM)

• Search for places by category within a bounding box

• Suggest optimal meeting points for multiple people

• Explore areas and run neighborhood livability analysis

• Find schools, EV charging stations and parking facilities

• Map layers: fetch rendered map tiles (standard, transport, cycle, …) as images, and render composed, annotated map images (markers + routes/lines drawn on top)

• Public transport layer: stops, stations and transit route lines (bus, tram, train, subway, light rail, ferry), plus full network topology — lines with ordered stations, interchanges and connections

• Energy layer: power lines, underground cables, substations and transformers (with voltage filtering), electricity production facilities (power plants and generators, filterable by source and output), full power-line tracing and territorial grid summaries

• Trip tooling: travel time/distance matrices (N×M) and POI search along a route corridor

Agent ergonomics

• Locations as text everywhere: every point-based tool accepts either coordinates or a location string (place name/address), geocoded automatically — no need to chain geocode_address first. Route-style tools take from_location/to_location (or home_location/work_location), and search_category accepts a named area.

• Compact responses by default: Overpass-based tools omit raw OSM tags unless verbose: true is passed, keeping token usage low in agent loops.

Related MCP server: OpenStreetMap MCP Server

Installation

Requires Node.js ≥ 18.17.

git clone https://github.com/visuelconcept/myosm-mcp-server.git
cd myosm-mcp-server
npm install
npm run build

Running the server

The server supports both MCP transports; stdio is the default.

stdio (local MCP hosts: Claude Desktop, Claude Code, Cursor, Windsurf, …)

{
  "mcpServers": {
    "myosm-mcp-server": {
      "command": "node",
      "args": ["/path/to/myosm-mcp-server/dist/index.js"]
    }
  }
}

With Claude Code:

claude mcp add myosm -- node /path/to/myosm-mcp-server/dist/index.js

Claude Desktop config file locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%/Claude/claude_desktop_config.json

Streamable HTTP (remote deployments, shared server, multiple clients)

Start the server in HTTP mode:

node dist/index.js --http --port 3000        # or: npm run start:http
# equivalent: MCP_TRANSPORT=http PORT=3000 node dist/index.js

The MCP endpoint is http://127.0.0.1:3000/mcp (sessions are managed per client via the Mcp-Session-Id header) and GET /health reports liveness for deployment probes.

Declare it in an MCP host by URL:

{
  "mcpServers": {
    "myosm-mcp-server": {
      "type": "http",
      "url": "http://127.0.0.1:3000/mcp"
    }
  }
}

With Claude Code:

claude mcp add --transport http myosm http://127.0.0.1:3000/mcp

CLI options: --http, --stdio, --host <address>, --port <number>, --help.

Security: the HTTP server binds to 127.0.0.1 by default and has no built-in authentication. To expose it (--host 0.0.0.0), put it behind a reverse proxy that handles TLS and auth, and set MCP_ALLOWED_HOSTS (comma-separated Host header values, e.g. MCP_ALLOWED_HOSTS=mcp.example.com) to enable DNS-rebinding protection.

Tools

Geocoding

Places & analysis

Routing

Map layers & transport

Energy

Resources

• location://place/{query} — information about a place by name (JSON)

• location://map/{style}/{z}/{x}/{y} — styled map tile at tile coordinates (PNG)

Configuration

All configuration is optional and done through environment variables:

Behind a corporate proxy, run Node with NODE_USE_ENV_PROXY=1 (Node ≥ 22.15) so fetch honors HTTPS_PROXY.

Staying inside the Nominatim rate limit

An LLM asked to locate a dozen places emits a dozen geocode_address calls in one turn, and MCP clients run them back to back — which is exactly what the public Nominatim answers with 429 Too Many Requests. The client therefore serializes geocoding requests with at least NOMINATIM_MIN_INTERVAL_MS between two of them, retries a 429 with back-off, and caches answers (concurrent lookups of the same place share a single request). Geocoding many places is correspondingly slower — roughly one second each — which is the price of the free service; point NOMINATIM_URL at a self-hosted instance and the spacing defaults to zero.

Fair use: by default the server talks to free, community-run services. Respect the Nominatim usage policy (max 1 req/s), the Overpass fair-use policy and the OSRM demo server policy. For production workloads, point the environment variables at your own instances. Note: the public OSRM demo server routes every profile with car data; self-host OSRM to get real bike/foot routing (the server already maps modes to the canonical driving/cycling/walking profiles).

Development

npm run build       # compile TypeScript to dist/
npm run watch       # recompile on change
npm test            # smoke + integration + HTTP transport tests (all offline)
SMOKE_LIVE=1 npm run smoke   # additionally exercise the real public OSM APIs
npm run inspector   # debug with the MCP Inspector

The integration and HTTP tests (test/integration.mjs, test/http.mjs) run the server against local mock implementations of Nominatim, OSRM, Overpass and the tile server, so they work without network access.

Differences from the Python original

Same 12 tools and 2 resources, plus:

• 10 new tools: get_map_tile, render_map, find_public_transport, get_transit_network, find_power_infrastructure, find_power_plants, trace_power_line, get_grid_summary, get_travel_time_matrix, search_along_route

• Locations accepted as free text everywhere (automatic geocoding) and compact responses by default (verbose: true for raw OSM tags)

• Streamable HTTP transport (--http) in addition to stdio, with per-client sessions and a /health endpoint

• Overpass queries use out center, so ways/relations (building-mapped schools, parking lots, …) return usable coordinates instead of being dropped

• search_category subcategory filtering uses a valid Overpass regex filter (the original generated invalid QL)

• Transport modes are mapped to canonical OSRM profiles (driving/cycling/walking)

• Endpoints are configurable via environment variables

• Turn-by-turn instructions are synthesized from OSRM maneuvers ("turn left onto …")

License

MIT — see LICENSE. Original Python implementation © open-streetmap-mcp contributors.