NeuronForm MCP server

NeuronForm turns OpenStreetMap data into print-ready, multi-colour 3D map models. Its public MCP server lets AI assistants such as Claude or ChatGPT design these models in a chat: find a place, create and change a design, show it as an image and, where the chat supports it, as a rotatable 3D view, and hand the user a link that opens exactly that design in the NeuronForm configurator.

This repository holds the documentation of the server, example briefs with the prompts that produce them, and the server.json for the MCP registry. It holds no code of the NeuronForm service.

German version: Deutsch. Website: neuronform.com/agents/en.

MCP address

https://neuronform.com/mcp

Agents never export print files and never buy. Exports and purchases happen only in the configurator, with the user's own NeuronForm account.

Related MCP server: OSM Edit MCP Server

Setup

Claude (claude.ai, desktop, and mobile apps)

Custom connectors are available on every Claude plan; the free plan allows one.

1. Open “Customize” › “Connectors”.

2. Choose “+ Add” and then “Add custom connector”.

3. Enter the name NeuronForm and the MCP address, and select “Continue”.

4. Review the detected authentication settings and select “Continue”.

5. Under “Authentication”, choose “No sign in”.

6. Select “Add”.

In a chat, switch the connector on under “+” › “Connectors”. In Team and Enterprise organizations an owner adds the connector under “Organization settings” › “Connectors” with “Add” › “Custom” › “Web”, and members connect it under “Customize” › “Connectors” with “Connect”. Source: Anthropic help article on custom connectors.

ChatGPT

Whether you can add your own MCP servers depends on the ChatGPT plan and, in a workspace, on its permissions and security settings.

1. Open ChatGPT Plugins in ChatGPT on the web.

2. Choose the plus button and then “Add custom MCP server”.

3. Enter the name NeuronForm and, optionally, a short description.

4. Under “Connection”, choose the public endpoint and enter the MCP address.

5. Choose “No authentication”, review the risk warning, and select “I understand and want to continue”.

6. Select “Create as a plugin”, review the tools the server lists, and install the plugin.

In a new conversation, type @ and choose NeuronForm. OpenAI changes these labels from time to time; its developer documentation holds the current steps.

Claude Code

claude mcp add --transport http neuronform https://neuronform.com/mcp

With --scope project the entry goes into the .mcp.json of the project.

Cursor

.cursor/mcp.json in the project, or ~/.cursor/mcp.json for every project:

{
  "mcpServers": {
    "neuronform": {
      "url": "https://neuronform.com/mcp"
    }
  }
}

VS Code

Run “MCP: Add Server” from the Command Palette, choose “HTTP (HTTP or Server-Sent Events)”, enter the MCP address and the server ID neuronform, and choose a destination:

• “Global” › “Copilot Global” writes ~/.copilot/mcp-config.json, or $COPILOT_HOME/mcp-config.json when COPILOT_HOME is set, for every workspace.

• .mcp.json at the root of the workspace shares the server with a team; the guided flow offers it with the experimental setting chat.mcp.workspaceRootConfig.enabled.

• .vscode/mcp.json in the workspace, which the guided flow labels deprecated.

The entry in .vscode/mcp.json:

{
  "servers": {
    "neuronform": {
      "type": "http",
      "url": "https://neuronform.com/mcp"
    }
  }
}

The portable files ~/.copilot/mcp-config.json and .mcp.json list the same entry under mcpServers instead of servers. Microsoft describes the destinations in its documentation of MCP servers in VS Code.

Other clients

Every client that connects to remote MCP servers over Streamable HTTP needs only the MCP address.

Tools

Every tool takes the optional argument lang (de or en, default en). It sets the language of the summaries, the notices, and the lead sentences and hints of error messages; the guide texts and the detail texts of the design validation are English. A successful call returns a text block with a short summary followed by the JSON result, and the same result as structuredContent; render_design returns the JPEG image before the text block, and the text block of get_design_guide holds the guide text. A failed call returns isError: true with an actionable message and, in structuredContent, the message and the list errors with one entry per refused field (field, message, hint); an unknown tool name is a JSON-RPC error. The design argument is the complete design JSON; its schema is the resource neuronform://schema/design.

Guide topics of get_design_guide: overview, footprint, tiles, colors_and_layers, terrain, frame, hanging, paths, texts_and_shapes, inserts, edge_labels, element_takeover, details, focus_scaling, custom_houses, limits.

Besides the fields of the brief, a design takes every field of the configurator through edit_design: among them texts and shapes sized in millimetres of the print or in metres on the map (topic texts_and_shapes), the generic details (roofs, windows, doors, and trees, with a facade style and a detail level, up to a map width of 3 km; topic details), focus scaling that magnifies one place and shrinks its surroundings toward the edge (topic focus_scaling), and up to five custom houses on free ground or in place of a building of the map (topic custom_houses).

A compass direction of edit_tiles selects the edge whose outward normal lies within 22.5° of it, so the directions follow the shape and the rotation: a hexagon without rotation offers ne, e, se, sw, w, and nw. validate_design and render_design list the free edges of every tile.

Brief

The brief of create_design and the changes of edit_design; in a brief only the place, or lat and lon, is required, and in changes an absent field keeps the design.

Workflow

1. search_places finds the place; create_design turns a brief into a complete design. The centre of the brief is the centre of tile 0.

2. edit_design changes brief fields and patches every other field; edit_tiles adds or removes tiles at the free edges that validate_design lists, and find_map_elements returns ready-to-insert roads, buildings, areas, and ground regions.

3. validate_design checks a changed design without generating it.

4. render_design shows the design as an image and an interactive 3D view.

5. share_design returns the link for the user.

Example prompts

• “Create a 3 km wide hexagonal relief model of Hallstatt with 1.5-fold exaggeration and show it to me in 3D.”

• “Make a square city map of central Vienna, 2 km wide, in two colours, with the edge text ‘Wien’, and give me a link to open it in the configurator.”

• “Build a wall map of Salzburg from three hexagonal tiles, add an integrated frame, and highlight the Getreidegasse in red.”

• “Show me the Großglockner as a round relief, 8 km wide, and compare 1.2-fold with 2-fold exaggeration.”

The folder examples holds the first three as JSON files. Each file names the prompt in English and German, the brief for create_design, the tool calls in their order, and the start link of its brief. The steps list the arguments besides brief and design: create_design receives the brief of the file, and every later call receives the current design.

Without an MCP client

An assistant takes the first of these three ways that it has:

1. The MCP tools of NeuronForm are connected: the assistant calls them.

2. The assistant runs code with internet access, with a code tool, a sandbox, or a terminal: it calls the MCP address directly over HTTP. Every request is one JSON-RPC POST answered with one JSON response; the server keeps no session, so no initialize is needed, and tools/list lists the tools with their schemas. create_design returns the design under result.structuredContent.design:

   curl -s https://neuronform.com/mcp \
     -H 'content-type: application/json' \
     -H 'accept: application/json, text/event-stream' \
     -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"create_design","arguments":{"brief":{"place":"Hallstatt","width":3000,"shape":"hexagon","relief":true,"exaggeration":1.5}}}}'

   The assistant passes that design as DESIGN to the next calls: render_design returns a JPEG of the model as base64 image content and a link to the 3D viewer under structuredContent.viewer, valid for one hour; share_design returns the link under structuredContent.url that opens exactly this design in the configurator, valid for 90 days.

   {"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"render_design","arguments":{"design":DESIGN}}}
   {"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"share_design","arguments":{"design":DESIGN}}}

3. The assistant only reads web pages: it writes a start link with the place and the settings (next section) and gives it to the person. Opening the link builds the real model in the configurator with its 3D preview; the assistant says that the model appears when the link is opened and presents no picture of its own as a NeuronForm model.

Start links for chats without MCP

A chat without MCP can hand over a design as a start link of the configurator:

https://neuronform.com/map/app?place=Hallstatt&width=3000&shape=hexagon&relief=true&exaggeration=1.5

With colours of its own, here sand ground, teal green areas, navy water, and coral roads:

https://neuronform.com/map/app?place=Ternitz&width=1800&shape=hexagon&colors=four&color_ground=f4e3c1&color_green=2a9d8f&color_water=1e3a8a&color_traffic=e76f51

The query parameters are the fields of the brief: place, lat, lon, width, shape, sides, rotation, tiles, model_width, colors, relief, exaggeration, frame, and label; colors takes a template name only. A parameter color_<key> sets the colour RRGGBB, without #, of one key of the colors object of the brief: a group ground, green, water, or traffic, a layer kind such as buildings or sea, frame, or label; these colours apply after the template. The language of the notices is that of the configurator. Only place, or lat and lon, is required. Values are URL-encoded, such as %20 for a space. Numbers are decimal numbers, and relief is true or false. The configurator builds the design with the compiler of create_design and names every value it adjusted; a link with a malformed value, or with a place the place search does not find, opens the configurator without the design and names the reason.

Limits and fair use

Use is free of charge. Bulk retrieval and circumventing the limits are not permitted; see the terms below.

Maintenance

While the operator puts the configurator, for visitors or for accounts, or the whole application into maintenance, the MCP address answers every request with HTTP 503 instead of a JSON-RPC response, so no tool runs and MCP clients report a failed request; the downloads of the 3D viewer answer the same way. The answer carries:

• the header Retry-After with the seconds until the expected end of the maintenance, at least 60 and at most 86,400, or 300 without an expected end;

• a problem JSON body of the type tag:neuronform.com,2026:maintenance with the code maintenance.configurator or maintenance.site and the object maintenance with scope, the messages of the operator message_de and message_en, and the expected end expected_end (RFC 3339, UTC), each null when the operator set none.

The switches in force are public at neuronform.com/map/api/v1/maintenance. Retry after the time that Retry-After names.

Privacy and terms

A request contains the tool arguments, such as the place and the design, and the IP address of the calling system: the platform of the chat for Claude or ChatGPT, the user's own computer for Claude Code, Cursor, or VS Code. When a client connects, it names its client software and version, which NeuronForm counts for its usage statistics. NeuronForm receives no account data and no other content of the chat. Place searches are answered by the place directory of the NeuronForm server, which it builds from OpenStreetMap data; the search terms reach no further service. The providers of the assistants are responsible for the chat contents on their platforms.

• Privacy policy: English, Deutsch

• Terms, section on AI agents: English, Deutsch

Map data and attribution

The service names its data sources from one data source registry. Every result of render_design carries, in the field attribution, the attribution lines of the data sources of its model in the language of the call, and its image carries the lines of the map data and terrain sources; the sea-floor depths are named exactly when the model contains them. Keep the attribution when you show or publish images of a design.

The registry lists every source with its role, attribution line, licence, the notice its licence requires, and its citation:

• as JSON: neuronform.com/map/api/v1/data-sources

• as a page: neuronform.com/licenses/en

MCP registry

server.json describes the server for the official MCP registry under the name com.neuronform/mcp, a namespace that the registry verifies against the domain neuronform.com. It follows the schema https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json. The registry lists the server, version 0.1.0, at registry.modelcontextprotocol.io/v0/servers?search=com.neuronform. It was published with mcp-publisher login http --domain neuronform.com, whose Ed25519 public key neuronform.com serves at /.well-known/mcp-registry-auth.

Contact

Questions and feedback: office@neuronform.com.

License

The documentation and the examples of this repository are licensed under the MIT License. The license does not cover the NeuronForm service or its results, such as images, 3D views, and share links; the terms of NeuronForm govern them.

Deutsch

NeuronForm macht aus OpenStreetMap-Daten druckfertige, mehrfarbige 3D-Kartenmodelle. Über den öffentlichen MCP-Server können KI-Assistenten wie Claude oder ChatGPT diese Modelle im Chat gestalten:

• einen Ort finden;

• einen Entwurf erstellen und ändern;

• ihn als Bild und, wo der Chat das unterstützt, als drehbare 3D-Ansicht zeigen;

• einen Link übergeben, der genau diesen Entwurf im Konfigurator öffnet.

Exportieren und kaufen können Agenten nicht; das geht nur im Konfigurator mit einem eigenen Konto.

• MCP-Adresse: https://neuronform.com/mcp (Streamable HTTP, Anfragen bis 2 MiB, ohne Konto, ohne API-Schlüssel, kostenlos).

• Einrichtung: In Claude unter „Customize“ › „Connectors“ › „+ Add“ › „Add custom connector“, Authentifizierung „No sign in“. In Claude Code mit claude mcp add --transport http neuronform https://neuronform.com/mcp. Für ChatGPT, Cursor und VS Code siehe die Abschnitte oben oder die Seite KI-Integration.

• Werkzeuge: search_places, create_design, edit_design, get_design_guide, find_map_elements, edit_tiles, validate_design, render_design, share_design, list_templates. Jedes Werkzeug nimmt den optionalen Parameter lang (de oder en, Standard en) für Zusammenfassungen, Hinweise und Fehlermeldungen.

• Ohne MCP-Client: Ein Assistent mit Code-Werkzeug, Sandbox oder Terminal ruft die MCP-Adresse direkt per HTTP auf: jede Anfrage ist ein JSON-RPC-POST mit einer JSON-Antwort, ohne Sitzung und ohne initialize (Beispiel mit curl im Abschnitt „Without an MCP client“). Der Accept-Header darf application/json, application/json, text/event-stream oder */* nennen oder fehlen; nur text/event-stream allein beantwortet der Server mit HTTP 406.

• Startlinks ohne MCP: https://neuronform.com/map/app?place=Hallstatt&width=3000&shape=hexagon&relief=true&exaggeration=1.5; eigene Farben setzen Parameter color_<Schlüssel> mit RRGGBB ohne # für eine Gruppe (ground, green, water, traffic), eine Ebene wie buildings oder sea, frame oder label, zum Beispiel https://neuronform.com/map/app?place=Ternitz&width=1800&shape=hexagon&colors=four&color_ground=f4e3c1&color_green=2a9d8f&color_water=1e3a8a&color_traffic=e76f51

• Grenzen:

  • Es gelten die Grenzen eines Besuchs ohne Konto (bis zu 3 Kacheln und 20 eigene Linien).

  • Generische Details (Dächer, Fenster, Türen und Bäume) gibt es bis 3 km Kartenbreite; einen breiteren Entwurf mit eingeschalteten Details lehnt der Server ab. Eigene Häuser gibt es bis zu fünf je Entwurf mit einem Detailbudget von 2 Millionen Dreiecken je Haus.

  • Alle Agenten zusammen können pro Stunde 2.000 Modelle neu berechnen lassen; Ansichten eines noch zwischengespeicherten Modells zählen nicht.

  • Besucherinnen und Besucher der Website haben Vorrang.

  • Ortssuchen, Entwurfsprüfungen und Bilder laufen in einer begrenzten Kapazität, die die Website mitnutzt; ist sie ausgeschöpft, lehnt der Server weitere Aufrufe dieser Art vorübergehend ab.

  • Modelle mit mehr als etwa einer Million Dreiecken lehnt der Server für Agenten ab.

  • Ein Modell liest höchstens 2 GiB (2048 MiB) dekodierte Kartendaten; einen Kartenausschnitt mit mehr Kartendaten lehnt der Server ab, bis ein kleinerer Ausschnitt oder weniger Details gewählt sind.

• Wartung: Solange der Betreiber den Konfigurator oder die ganze Anwendung in Wartung schaltet, beantwortet die MCP-Adresse jede Anfrage mit HTTP 503 statt mit einer JSON-RPC-Antwort. Die Antwort nennt in Retry-After die Sekunden bis zum erwarteten Ende (ohne Ende 300) und im Objekt maintenance die Nachricht und das erwartete Ende; den Stand zeigt neuronform.com/map/api/v1/maintenance.

• Datenschutz und Bedingungen: Datenschutzerklärung, Allgemeine Geschäftsbedingungen, Abschnitt „Nutzung durch KI-Agenten“.

• Kartendaten und Quellen: Jedes Ergebnis von render_design trägt im Feld attribution die Namensnennung der Datenquellen seines Modells, das Bild die der Karten- und Geländedaten. Alle Quellen mit Lizenz, Hinweis und Zitat listet das Quellenverzeichnis des Dienstes: als JSON und auf der Seite Lizenzen und Attribution.