
TomTom Maps MCP Server
MCPTomTom's official MCP: geocoding and reverse geocoding, fuzzy and nearby search, single and multi-stop routing, reachable range, live traffic incidents, and static or dynamic maps. Needs a TomTom developer account and API key; use the hosted server (public preview) or run locally via npx.
TomTom Maps MCP Server
The TomTom Maps MCP Server simplifies geospatial development by providing seamless access to TomTom’s location services, including search, routing, traffic and interactive maps. It enables easy integration of precise and accurate geolocation data into AI workflows and development environments.
Demo

Table of Contents
- Demo
- Security Notice
- Remote MCP Server (No Installation Required)
- Quick Start
- Integration Guides
- Available Tools
- Debug UI
- Local Development
- Troubleshooting
- Contributing & Feedback
- Security
- License
Remote MCP Server (No Installation Required)
Public Preview — The TomTom Maps Remote MCP Server is currently in public preview.
The easiest way to get started is to connect directly to TomTom's hosted MCP Server — no Node.js, Docker, or local setup needed.
Endpoint:
https://mcp.tomtom.com/maps
Prerequisites:
- A valid TomTom API key with MCP Server access enabled (see API Key Management)
Generic MCP Client Configuration
Add the following to your MCP client configuration:
{
"mcpServers": {
"tomtom-mcp": {
"type": "http",
"url": "https://mcp.tomtom.com/maps",
"headers": {
"tomtom-api-key": "your_api_key_here"
}
}
}
}
VS Code (GitHub Copilot)
Create or edit .vscode/mcp.json in your workspace:
{
"servers": {
"tomtom-mcp": {
"type": "http",
"url": "https://mcp.tomtom.com/maps",
"headers": {
"tomtom-api-key": "your_api_key_here"
}
}
}
}
Claude Desktop
The quickest option is to install the pre-built extension — see the Claude Desktop Setup guide for details.
Alternatively, configure Claude Desktop to use the remote server directly by editing your configuration file:
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json
{
"mcpServers": {
"tomtom-mcp": {
"type": "http",
"url": "https://mcp.tomtom.com/maps",
"headers": {
"tomtom-api-key": "your_api_key_here"
}
}
}
}
Note: If your MCP client does not support remote HTTP connections with custom headers, use the local setup instead.
Security Notice
Keeping local deployments of the TomTom Maps MCP Server up-to-date is the responsibility of the MCP client/operator. TomTom publishes updates to address known vulnerabilities, but failing to apply updates, patches, or recommended security configurations to your local instance may expose it to known vulnerabilities.
Quick Start
Prerequisites
- Node.js 22.x
- TomTom API key
How to obtain a TomTom API key:
- Create a developer account on TomTom Developer Portal and Sign-in
- Go to API & SDK Keys in the left-hand menu.
- Click the red Create Key button.
- Select all available APIs to ensure full access, assign a name to your key, and click Create.
For more details, visit the TomTom API Key Management Documentation.
Installation
npm install @tomtom-org/tomtom-mcp@latest
# or run directly without installing
npx @tomtom-org/tomtom-mcp@latest
Configuration
Set your TomTom API key using one of the following methods:
# Option 1: Use a .env file (recommended)
echo "TOMTOM_API_KEY=your_api_key" > .env
# Option 2: Environment variable
export TOMTOM_API_KEY=your_api_key
# Option 3: Pass as CLI argument
TOMTOM_API_KEY=your_api_key npx @tomtom-org/tomtom-mcp@latest
Environment Variables
| Variable | Description | Default |
|---|---|---|
TOMTOM_API_KEY | Your TomTom API key | - |
PORT | Port for the HTTP server | 3000 |
LOG_LEVEL | Logging level: debug, info, warn, or error. Use debug for local development to see all logs | info |
MCP_APPS | Stdio only. false gives the client only the data tools, without maps; true gives every tool. Unset, the client's initialize decides. Over HTTP, use ?apps= on the MCP URL (see Clients without MCP apps) | - |
Usage
Stdio Mode (Default - for AI assistants like Claude):
# Start MCP server via stdio
npx @tomtom-org/tomtom-mcp@latest
HTTP Mode (for web applications and API integration):
pnpm run build # Build first (required)
pnpm run start:http
# or run the built binary directly
node bin/tomtom-mcp-http.js
When running in HTTP mode, you need to include your API key in the tomtom-api-key header:
tomtom-api-key: <API_KEY>
For example, to make a request using curl:
curl --location 'http://localhost:3000/mcp' \
--header 'Accept: application/json,text/event-stream' \
--header 'tomtom-api-key: <API KEY>' \
--header 'Content-Type: application/json' \
--data '{
"method": "tools/call",
"params": {
"name": "tomtom-geocode",
"arguments": {
"query": "Amsterdam Central Station"
}
},
"jsonrpc": "2.0",
"id": 24
}'
The Docker setup is also configured to use this HTTP mode with the same authentication method.
Docker Mode (recommended):
# Option 1: Using docker run directly
docker run -p 3000:3000 ghcr.io/tomtom-international/tomtom-maps-mcp:latest
# Option 2: Using Docker Compose (recommended for development)
# Clone the repository first
git clone https://github.com/tomtom-international/tomtom-maps-mcp.git
cd tomtom-maps-mcp
# Start the service
docker compose up
Both Docker options run the server in HTTP mode. Pass your API key via the tomtom-api-key header as shown in the HTTP Mode curl example above.
Integration Guides
TomTom Maps MCP Server can be easily integrated into various AI development environments and tools.
These guides help you integrate the MCP server with your tools and environments:
- Claude Desktop Setup - Instructions for configuring Claude Desktop to work with TomTom Maps MCP server
- VS Code Setup - Setting up a development environment in Visual Studio Code
- Cursor AI Integration - Guide for integrating TomTom Maps MCP server with Cursor AI
- Windsurf Integration - Instructions for configuring Windsurf to use TomTom Maps MCP server
- Smolagents Integration - Example showing how to connect Smolagents AI agents to TomTom Maps MCP server.
Available Tools
How dynamic map tool works
The dynamic map tool renders nothing server-side. It resolves the request into map state — the basemap style to load, the viewport to open on, and GeoJSON sources and layers for the markers, routes and polygons requested — calculating any routePlans through the Routing API along the way.
That state is cached and the tool returns its viz_id. The MCP app fetches it with the app-only tomtom-get-viz-data tool and draws the map client-side, so panning, zooming and clicking work on a live map.
Because the map is drawn by the app, the visual requires an MCP client that supports MCP apps. A client the server knows renders no MCP apps, such as Claude Code, Gemini CLI, Cline or Zed, does not get tomtom-dynamic-map, tomtom-data-viz, the app-only tools or show_ui at all (see Clients without MCP apps). Any other client without MCP apps receives a JSON summary of what the map shows: its view, markers, routes (distance, travel time, traffic delay) and areas.
Clients without MCP apps
When a client connects, its initialize request decides what it gets, over stdio and HTTP alike:
| The client | Gets |
|---|---|
advertises the MCP Apps extension (io.modelcontextprotocol/ui with text/html;profile=mcp-app): Claude (web, desktop, mobile), ChatGPT, VS Code, Cursor, Goose Desktop | every tool, with maps |
is known to render no MCP apps, by its clientInfo.name: Claude Code, Gemini CLI, Cline, Zed, Goose CLI, Continue, Amazon Q CLI, opencode, Roo Code | the data tools only, without show_ui, app resources or the _meta the apps read |
| neither | every tool, as before: some clients render apps without advertising the extension |
The HTTP server keeps no state, so its initialize response returns an Mcp-Session-Id of the form <apps>.<client>.<uuid> recording this classification and the client's clientInfo.name, base64url-encoded (or unknown without one), and the client sends it with every later request. The name is also reported as client in the TomTom-Upstream-Metadata response header, next to auth_method, for the gateway's analytics.
A client can also choose for itself, which is useful for a headless agent or a client in the last row. Over HTTP, add ?apps=false to the MCP URL (for example https://mcp.tomtom.com/maps?apps=false) for the data tools only, or ?apps=true for every tool. Over stdio, set MCP_APPS=false or MCP_APPS=true in the server's environment.
The map apps load MapLibre GL JS from https://cdn.jsdelivr.net rather than bundling it, so the client must be able to reach jsDelivr as well as api.tomtom.com. Both are declared in the apps' content security policy. MapLibre's worker is bundled into each app instead, because ChatGPT's sandbox allows workers only from blob: URLs.
References:
- TomTom Orbis Maps style: https://developer.tomtom.com/map-display-api/documentation/tomtom-orbis-maps/vector-style
Getting geometry out of a tool response
Every tool that returns TomTom data accepts a response_detail parameter (all but tomtom-poi-categories, tomtom-dynamic-map and tomtom-data-viz). The four tools that return geometry (tomtom-routing, tomtom-traffic, tomtom-area-search and tomtom-search-along-route) accept three values; the others accept compact and full.
| Value | Returns |
|---|---|
compact (default) | Essential fields and the point coordinates of a place. No geometry: route lines, traffic incident locations and search areas are omitted. |
geometry | compact, plus a geometry key holding that geometry as a GeoJSON FeatureCollection. |
full | The raw API response: lossless, in the API's own shape, and many times larger. |
The default is tuned for conversational use, where a route line would consume most of a model's context for no benefit. If you are building on top of the server and need the coordinates themselves, to draw the result on your own map or run your own analysis, request response_detail: "geometry". Use full only when you need fields that compact drops, or the exact line.
For an Amsterdam-to-Berlin route, geometry is about 22 KB: the 8,000-point line is simplified to 1,000 vertices, at most 13 m from the original. full is about 210 KB.
The geometry FeatureCollection
{
"type": "FeatureCollection",
"features": [
{
"type": "Feature",
"properties": { "summary": { "lengthInMeters": 663425, "travelTimeInSeconds": 24453 } }
}
],
"geometry": {
"type": "FeatureCollection",
"features": [
{
"type": "Feature",
"geometry": { "type": "LineString", "coordinates": [[4.90413, 52.36761], [4.90419, 52.36755]] },
"properties": {
"route": 0,
"simplification": { "original_points": 8151, "points": 1000, "max_error_m": 13 }
}
}
]
}
}
-
Coordinates follow RFC 7946:
[longitude, latitude], rounded to 5 decimal places (about 1.1 m). Polygon rings are closed. -
One feature per item. Its
propertieshold only a join key giving the item's position in the rest of the response. A join key is valid within one response only; don't store it as an identifier.Tool Features propertiesRouting One LineStringper route{"route": 0}Traffic One PointorLineStringper incident, as the API returns it{"incident": 12}, matchingincidents[12]Area search The search boundary Polygon{"boundary": "circle"},"polygon"or"boundingBox"Search along route The route LineString{"route": 0} -
At most 1,000 vertices per feature. Longer lines are simplified, and the feature then carries
simplification: the original and returned vertex counts, andmax_error_m, an upper bound in whole metres on the distance between a dropped vertex and the returned line, including the shift from rounding coordinates to 5 decimals. A long route is accurate at the zoom that shows all of it, but visibly approximate when zoomed in; ifmax_error_mis too large for your use, requestfull. A polygon that would cross itself after simplification keeps more vertices instead, so it can exceed 1,000. -
No vertex indexes. Route sections and legs point into the API's original line, which a simplified line no longer matches, so
geometryresponses dropstartPointIndex,endPointIndexandpointIndex.
The design is recorded in docs/adr/.
Note: Hosts that support MCP Apps render the interactive map widget from the untrimmed response regardless of this setting, so
compactloses nothing visually. Theshow_uiparameter requests that widget and is ignored by hosts that cannot render it; it is not a way to obtain coordinates. Clients known to render no MCP apps don't get the parameter.
Debug UI
A built-in debug UI lets you visually test MCP tools and their interactive map widgets without needing an AI client.
Quick Start
pnpm run ui
This starts both the MCP HTTP server (port 3000) and the debug UI host (port 8080). Open http://localhost:8080 in your browser.
Features
- Tool browser — searchable sidebar listing all available tools, with icons distinguishing map-enabled tools from plain tools
- Pre-filled examples — each tool loads with example parameters (including
show_ui: truefor map widgets) - Live map widgets — tools with UI resources render interactive TomTom maps directly in the browser
- Response metadata — latency, payload size, estimated token count, content parts, and timestamps for every call
- Dark / light mode — toggle with the theme button or follows system preference
- Keyboard shortcuts —
Cmd+Enterto run,Cmd+Kto search tools
Requirements
- The MCP server must be running in HTTP mode (handled automatically by
pnpm run ui) - A valid
TOMTOM_API_KEYin your.envfile
Building the UI separately
The UI host is a workspace package (tomtom-mcp-app-host in ui/), so the root pnpm install already installed its dependencies.
pnpm run ui:build # Build the UI
pnpm --filter tomtom-mcp-app-host start # Start only the UI host (assumes MCP server is already running)
Local Development
This project uses pnpm (
>=11) as its package manager. Install it withnpm install -g pnpmorcorepack enable. Linting and formatting are handled by Biome.
Setup
git clone https://github.com/tomtom-international/tomtom-maps-mcp.git
cd tomtom-maps-mcp
pnpm install
cp .env.example .env # Add your API key in .env
pnpm run build # Build TypeScript files
node ./bin/tomtom-mcp.js # Start the MCP server
Testing
pnpm run build # Build TypeScript
pnpm test # Run all tests
pnpm run test:all # All tests (unit + stdio + http)
Testing Requirements
⚠️ Important: The tests require a valid API key in .env: the service and tool tests make real API calls, which consume your API quota.
OAuth end-to-end test
pnpm run test:e2e:oauth connects to the deployed MCP server with the MCP SDK client, the library the connectors are built on, and follows the same path a connector does: the server's 401 names the authorization server, the SDK discovers it and opens the authorize request, a browser picks the sign-in tenant and logs the test user in, the SDK redeems the code, and the session calls a tool with the token. It runs for every account and connector of a target, with the same connectors on both targets, so a change on dev is checked the way real users will meet it on prod.
| Target | MCP server | Authorization server it must advertise | Connectors | Last step |
|---|---|---|---|---|
prod (default) | mcp.tomtom.com/maps | oauth.my.tomtom.com | ChatGPT, Claude | geocode returns a result |
dev | mcp-dev.api-system.tomtom.com/maps | staging.oauth.my.tomtom.com | ChatGPT, Claude | geocode returns a key rejection |
local | localhost:3000/mcp, started by you | staging.oauth.my.tomtom.com | the test client dev serves | geocode returns a key rejection |
The authorization server is discovered from the MCP server, never configured in the test; the test only asserts that the discovered one is the expected one, so a deployment pointing at the wrong ULS instance fails in the first step with that message. The connectors are the registered redirect URIs of ChatGPT and Claude; the sign-in tenant is chosen on the ULS page the way a user does, with no tenant shortcut.
On dev the last step proves the key exchange and not the tool result: most API services behind ULS staging are not fully functional, so the geocode call comes back with a key rejection, and that rejection is what the test asserts.
Select the target with OAUTH_E2E_TARGET. Accounts come from the environment, .env locally. A missing pair skips that account locally and fails the run in CI, so a misconfigured secret can never pass.
E2E_B2B_EMAIL,E2E_B2B_PASSWORD: workforce user,prodonly. The workforce tenant holds no resource app for the dev origin yet, so that login cannot succeed ondev.E2E_B2C_PROD_EMAIL,E2E_B2C_PROD_PASSWORD: external user onprod.E2E_B2C_DEV_EMAIL,E2E_B2C_DEV_PASSWORD: external user ondevandlocal.
On prod and dev the test needs no server running and starts none. The browser resolves the connectors' redirect hosts to 127.0.0.1, so a captured authorization code never reaches a third party.
local checks the OAuth path of a build on this machine without a dev deployment. Start the server with the env block of the dev deployment in tomtom-mcp-infra (.kaap/helm/dev/app.yaml), leaving out MCP_BASE_URL and MCP_BASE_URL_PATH so it advertises itself as localhost, then run the test against it:
AUTHORIZATION_SERVER_URL=... ULS_TOKEN_ENDPOINT=... ULS_CLIENT_ID=... CIAM_TENANT_ID=... CIAM_DOMAIN=... WORKFORCE_TENANT_ID=... pnpm run dev:http
OAUTH_E2E_TARGET=local pnpm run test:e2e:oauth
ChatGPT and Claude cannot reach a local server, so local logs in through the test client that dev serves at /maps/inspector-client.json, registered on ULS staging with a loopback redirect. That is why local depends on dev being up, and why the workforce account is left out: the dev origin has no workforce resource app.
The OAuth end-to-end workflow runs it in CI and posts to Slack only when the outcome changes:
- on a schedule against
prod; - on every dev deployment announced by
tomtom-mcp-infra, after waiting for the rollout; - on demand for either target.
Project Structure
src/
├── apps/ # MCP App UI resources
├── handlers/ # Request handlers
├── schemas/ # Validation schemas
├── services/ # TomTom API wrappers
├── tools/ # MCP tool definitions
├── types/ # TypeScript type definitions
├── utils/ # Utilities
├── createServer.ts # MCP Server creation logic
├── index.ts # Main entry point (stdio)
└── indexHttp.ts # HTTP server entry point
Troubleshooting
API Key Issues
echo $TOMTOM_API_KEY # Check if set
Test Failures
ls -la .env # Verify .env exists
cat .env # Check API key
Build Issues
pnpm run build # Rebuild
pnpm store prune # Clear cache
Forbidden (403) Errors
If you see an error stating "missing permissions", it means your API key does not have access to the TomTom Orbis Maps services, which back all of this server's tools.
Note: TomTom Orbis Maps are currently in Public Preview. They may not be available on all developer accounts by default.
How to troubleshoot:
- Log in to the TomTom Developer Portal.
- Ensure all available products are selected for your API key.
- If you still see 403 errors, your account may not yet have access to the Orbis preview — request access via the developer portal.
Contributing & Feedback
We welcome contributions to the TomTom Maps MCP Server! Please see CONTRIBUTING.md for details on how to submit pull requests, report issues, and suggest improvements.
All contributions must adhere to our Code of Conduct and be signed-off according to the Developer Certificate of Origin (DCO).
Open issues on the GitHub repo
Security
Please see our Security Policy for information on reporting security vulnerabilities and our security practices.
License
This project is licensed under the Apache License 2.0 - see the LICENSE.md file for details.
Copyright (C) 2025 TomTom Navigation B.V.
Похожие ресурсы

mcp-google-map
MCPGoogle Maps for agents: place search and details, forward and reverse geocoding, routes, distance matrices, multi-stop optimization, weather and static maps, plus trip planning and local SEO ranking skills. Needs Node.js and a billing-enabled Google Maps Platform API key; calls may cost money.
cablate/mcp-google-map470
QGIS MCP
MCPConnects the QGIS desktop GIS to Claude: create and save projects, load vector and raster layers, run Processing toolbox algorithms, export maps and execute PyQGIS code through chat. Needs QGIS 3 and uv; download the repo, copy the plugin into QGIS's plugin folder and start the server.
jjsantos01/qgis_mcp1.1k
QGIS MCP (nkarasiak)
MCPInstall the QGIS MCP plugin from the QGIS plugin repository and agents get 125 tools to manage layers, edit features, run processing algorithms, style maps, build layouts and atlases, and run SQL across layers; supports QGIS 3.28 to 4.x. Needs QGIS and uv; no repo download required.
nkarasiak/qgis-mcp364
GIS MCP Server
MCPWraps Shapely, GeoPandas, PyProj, Rasterio, PySAL and other GIS libraries so agents can run geometry and overlay analysis, reproject coordinates, process rasters (e.g. NDVI), do spatial statistics and make maps, and fetch boundaries, land cover and satellite imagery. Needs Python 3.10+ and uv.
mahdin75/gis-mcp198
GeoServer MCP Server
MCPConnects to GeoServer's REST API so agents can manage workspaces, data stores, layers and styles, run CQL spatial queries, and access features and render maps via WMS/WFS, for publishing and maintaining map services. Needs a running GeoServer with an admin account, Python 3.10+ and uv.
mahdin75/geoserver-mcp95
OpenCage Geocoding MCP Server
MCPOpenCage's official MCP: geocode addresses and place names, reverse geocode coordinates (with admin area, time zone and more), and check API quota. For small lookups; use its CLI for batches. Needs an OpenCage API key (free trial), Node.js 20+, and manual setup from a repo clone.
OpenCageData/opencage-geocoding-mcp19