Copy page

MCP server | Rijwind docs

MCP server

Rijwind ships a hosted MCP (Model Context Protocol) server, so AI assistants and coding agents can use the platform directly: geocode addresses, calculate routes and matrices, compute isochrones, match GPS traces, optimize trips, and render static map images — all as native tools.

Code
https://mcp.rijwind.com

The server speaks streamable HTTP and supports two ways to authenticate:

• API key — set your regular Rijwind key as a Bearer header. For coding agents (Claude Code, Codex, Cursor).
• OAuth sign-in — one-click connect with a browser login, no key to copy. For the Claude and ChatGPT apps (custom connectors).

MCP access is included in every paid plan. The Personal plan (€5/month) is the cheapest way in. Free-tier accounts get 403 mcp_not_enabled.

Connect your client

For the API-key clients below you need an API key. MCP calls run server-side, so use a key without an origin whitelist (an origin-restricted key is rejected with 403 origin_required).

Claude Code

Code
claude mcp add --transport http rijwind https://mcp.rijwind.com \
  --header "Authorization: Bearer rw_live_•••YOUR_KEY•••"

This covers the desktop app too: it shares its configuration with the CLI, so a server added with claude mcp add (or via .mcp.json) appears there automatically. The in-app Connectors menu only lists pre-built integrations — custom servers are added via the command or config file.

To share the server with your team instead, commit it to your project's .mcp.json:

Code
{
  "mcpServers": {
    "rijwind": {
      "type": "http",
      "url": "https://mcp.rijwind.com",
      "headers": {
        "Authorization": "Bearer rw_live_•••YOUR_KEY•••"
      }
    }
  }
}

Codex

Codex configures remote MCP servers in ~/.codex/config.toml (the codex mcp add command is for local stdio servers only). It reads the key from an environment variable, so the secret stays out of the config file:

Code
[mcp_servers.rijwind]
url = "https://mcp.rijwind.com"
bearer_token_env_var = "RIJWIND_API_KEY"

Then export the key in your shell profile, so it's set when Codex launches:

Code
export RIJWIND_API_KEY="rw_live_•••YOUR_KEY•••"

The Codex IDE extension shares this configuration with the CLI: open the Codex side panel → gear icon → MCP settings → Open config.toml to edit the same file from your editor.

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global). Cursor interpolates ${env:…} from your shell environment — recommended over hardcoding the key, especially in a project file you might commit:

Code
{
  "mcpServers": {
    "rijwind": {
      "url": "https://mcp.rijwind.com",
      "headers": {
        "Authorization": "Bearer ${env:RIJWIND_API_KEY}"
      }
    }
  }
}

A literal "Bearer rw_live_•••YOUR_KEY•••" value works too.

Once added, the server shows up under Cursor Settings → MCP, where you can toggle it on and off. Cursor's one-click marketplace installs only cover its own directory — custom servers like this one are added via mcp.json.

Claude & ChatGPT apps (OAuth, no key needed)

The Claude apps (web and desktop) and ChatGPT connect via custom connectors with a browser sign-in — no API key to copy:

1. Add a custom connector and enter https://mcp.rijwind.com as the URL (in Claude: Settings → Connectors → Add custom connector).
2. Your browser opens a Rijwind sign-in: log in with your dashboard account and approve the connection.
3. Done — the assistant can now use the Rijwind tools.

Behind the scenes the connection bills through a dedicated "MCP Connector" API key that is created automatically in your dashboard on first use, drawing from the same monthly unit pool. Revoking that key simply rotates it — to actually disconnect, remove the connector in the client (and the key from your dashboard if you like).

Other clients

Any MCP client that supports remote servers over streamable HTTP works with either auth mode:

Setting	Value
URL	https://mcp.rijwind.com
Header (API-key mode)	Authorization: Bearer rw_live_•••YOUR_KEY•••
OAuth	supported — discovery + dynamic client registration per the MCP spec

Tools

Tool	What it does	Cost (units)
geocode_forward	Search places / forward geocoding	1
geocode_reverse	Coordinates → nearest address	1
geocode_autocomplete	Complete a partial query	0.1
directions	Route along waypoints (car / bike / foot)	1
directions_matrix	Travel-time + distance matrix	sources × targets
isochrone	Reachable-area polygons	5 × locations × contours
map_matching	Snap a GPS trace to the road network	1
optimized_route	Optimal visiting order over stops	1 per location
static_map	Render a map image with overlays	1 (2 for @2x)
show_map	Interactive map with markers, routes, isochrones	1 per map session

These are the same endpoints as the REST API, with the same parameters and the same responses — the MCP server is a thin translation layer, not a separate product.

Maps in the conversation

Two tools draw maps, for different situations:

• static_map returns a real image — it renders inline in any client that displays image results (the Claude apps, Claude Code, Cursor) and is ideal for reports and quick previews.
• show_map opens an interactive map (pan, zoom, clickable markers) directly in the conversation, on the Rijwind basemap. It accepts markers, encoded route polylines (straight from directions), and GeoJSON (straight from isochrone). It uses the MCP Apps standard, supported by the Claude apps, ChatGPT, and VS Code, among others. Costs 1 unit per map session — panning, zooming, and tile refreshes within the session are free. In clients without MCP Apps support (e.g. terminal agents), the tool degrades to its structured JSON result; use static_map there.

Quotas and scopes

• Units come from your regular monthly pool. An MCP call costs exactly what the equivalent REST call costs — see Quotas & rate limits. There is no separate MCP meter.
• Key scopes filter the tool list. A key scoped to geocode only exposes the three geocoding tools; routing and static-map tools don't appear to the agent at all. An unscoped key exposes everything.
• Requests are rate-limited per key. Sustained agent workloads are fine; runaway loops get 429.

What the server won't do

The MCP server is deliberately data-plane only and read-only: no tool can change your subscription, spend money, or alter billing settings. Plan changes and payment settings live exclusively in the dashboard.