street-generator-mcp

An MCP server that lets Claude design and render urban street cross-sections from natural language, in the visual style of the Street Generator app, figures and all. It can also read a real street from an address via OpenStreetMap.

The idea

Claude does the open-ended reasoning (designing the street). The server does the deterministic work (drawing it).

The server holds no LLM, no API keys, and no secrets. It keeps the model's output small (a compact StreetConfig instead of thousands of tokens of hand-drawn SVG) and moves the heavy, exact rendering into code. The result is faster, cheaper, and reliable: every street is drawn by the same renderer, identical every time.

User: "a calm residential street, wide sidewalks, trees, protected bike lanes"
   |
   v
Claude -> StreetConfig (JSON), guided by the tool schema
   |
   |-- render_street(config, style)   -> server draws an illustrated SVG
   |-- build_share_url(config)        -> streetgenerator.com/?s=...
   |
   v
Claude shows the image and the link

Related MCP server: ladybug-tools-mcp

Tools

render_street(config, style?)

Renders a StreetConfig to an illustrated SVG cross-section (buildings as floor stacks, sidewalks with pedestrians, planting strips with trees, lanes with cars).

style (optional): colorMode "outline" (default) or "color", plus showLabels, showMeasurements, showFigures (all default true).

import_street_from_osm(street, houseNumber, city, postcode, country, style?)

Reads a real street from OpenStreetMap and renders it. Give a full address (street, house number, city, postcode, country). If any part is missing or unclear, Claude asks first. When several places match, the tool returns lettered candidates (A, B, C) so you can pick one, then it renders the chosen street.

Under the hood: geocode the address (Nominatim), read the nearest street's tags (Overpass), interpret and map them to a StreetConfig, then render. This tool needs internet access; the other two do not.

build_share_url(config)

Returns a streetgenerator.com link that opens the design in the live app, where you can keep editing and export it.

What you get back

Every render returns the cross-section as SVG (which you can save) together with a streetgenerator.com link that opens the same design in the app, where the app's own PNG, SVG, and JSON exports are available. Native PNG output from the server is on the roadmap (see Future work).

Examples

From simple to more complex. You type these to Claude; it calls the tools and shows the result.

Usage and external services

import_street_from_osm queries two free OpenStreetMap services: Nominatim (geocoding) and Overpass (map data). It is meant for occasional, interactive lookups, not bulk downloading. Please stay within their public usage policies:

• Nominatim: at most 1 request per second, an identifying User-Agent (already set by this server), results cached where possible, and no automated or bulk querying. See the Nominatim usage policy.

• Overpass: fair use only, roughly up to 10,000 requests and 1 GB of data per day per IP, with rate limiting under load and a 180 second per-request timeout. See the Overpass usage guidance.

For heavy or automated use, run your own Nominatim or Overpass instance instead of the public servers. Map data is copyright OpenStreetMap contributors, available under the ODbL; displayed results should attribute OpenStreetMap accordingly.

Install (Claude Desktop)

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

Then ask Claude to design or import a street and it calls the tools automatically.

Develop

npm install
npm test          # unit, mocked-network, and fidelity tests
npm run build     # compile to dist/ and copy figure assets
npm run dev       # run the server from source

How the rendering stays faithful to the app

The layout math, element data, figure art, and the OSM interpreter/mapper are ported value-for-value from the Street Generator app (the interpreter and mapper come with their original unit tests). Fidelity is checked with structural tests in test/.

Future work

• PNG output from the server (SVG to raster).

• validate_street against Berlin RASt rules (the engine already exists in the app).

• Nearest-way selection for OSM import (v1 takes the first street within a radius of the geocoded point).

• Figures for cycle and bus lanes once art exists.