Plateau Creative MCP

@yodolabs/plateau-creative-mcp

An MCP server that turns Project PLATEAU's 3D city data into scene-editing and glTF export tools for creative LLM agents.

English · 日本語

A real Shibuya block — loaded from PLATEAU, edited, and exported as one .glb by the export_glb tool, here rendered with three.js. Same geometry your LLM hands off to Blender / Unity / Unreal. Data © Project PLATEAU / MLIT (CC BY 4.0).

This server exposes tools. Orchestration is your LLM's job — point Claude / Cursor / any MCP client at this server alongside Blender MCP / Unity MCP / Unreal MCP and let it compose the workflow.

PLATEAU has an official MCP server for data access (catalog, specs, CityGML, attributes). This one fills the missing creation, scene-editing, and DCC export layer: load a Tokyo neighbourhood, delete every skyscraper, scale a block to 3× height, and emit a .glb ready for Blender / Unity / Unreal — all driven by tool calls.

flowchart LR

CLIENT["Claude / Cursor (any MCP client)"] -->|tool calls| MCP["plateau-creative-mcp"] MCP -->|download_area| DATA["plateau-bridge artifact index + bundles"] MCP -->|"load → filter → edit"| SCENE["scene_id (in-memory)"] MCP -->|export_glb| OUT[".glb + LICENSE.txt"] OUT --> DCC["Blender / Unity / Unreal"]

Status

v0.1.0 — first public release. 10 MCP tools (incl. on-demand artifact download — no data clone needed), polygon footprint extrusion with mesh merging, single-GLB export ≤ 1 km² (≈40% gltf-transform compression) with a scene_manifest 3D Tiles fallback for larger areas, three data-access modes (TS+duckdb artifact / Python subprocess / upstream MCP via JSON-RPC), Overpass POI linking, and an optional cross-MCP bridge to BlenderMCP. Vendor-neutral over MCP — Claude, Cursor, any client works.

Tools

When PLATEAU_UPSTREAM_ENABLED=true, the server also exposes 13 tools proxied from the official Project PLATEAU MCP server: plateau_spec_outline, plateau_spec_read, plateau_get_metadata, plateau_search_areas, plateau_get_area, plateau_search_datasets, plateau_get_dataset, plateau_list_dataset_types, plateau_citygml_get_attributes, plateau_citygml_get_features, plateau_citygml_get_geoid_height, plateau_get_citygml_files, plateau_explain_spatial_id. The LLM client sees 23 tools in a single catalog, with no need to wire two separate MCP servers.

Every tool returns { result, attribution_metadata }. export_glb writes asset.extras.attribution into the GLB and a LICENSE.txt next to it. You must keep that attribution in any derived video, image, or asset.

Quickstart

npm install

npm run build

There are two ways to get city data in place.

Option A — zero setup (recommended)

Don't clone the data pipeline at all. Out of the box the server reads the public plateau-bridge artifact index, so download_area works with no configuration — it fetches the prebuilt bundle for a city, verifies its sha256, and caches it locally:

export PLATEAU_AUTO_DOWNLOAD=true     # load_area auto-fetches a missing city (optional)

export PLATEAU_OUTPUT_DIR=./out npx @yodolabs/plateau-creative-mcp # stdio MCP server

Call download_area(city="shibuya") once (or set PLATEAU_AUTO_DOWNLOAD=true and just call load_area, which fetches a missing city first). Only buildings.parquet + manifest.json are extracted; bundles cache under PLATEAU_ARTIFACT_DIR (default ~/.cache/plateau-creative-mcp/artifacts), so each city downloads only once. Point PLATEAU_ARTIFACT_INDEX_URL at a different index to use your own mirror.

Downloaded bundles do not include 3D Tiles (they are gigabytes per city). Everything works from the parquet — load_area, all edits, and export_glb single_glb (real geometry, ≤ 1 km² / ≤ 5000 buildings). Only scene_manifest export loses its tile references for downloaded cities (it still emits the edit list, bbox, and attribution, with a tileset_note telling you to narrow to single_glb or run a local build). For full-geometry exports of larger areas, use Option B with a local build.

For Claude Desktop, add to claude_desktop_config.json:

{

"mcpServers": { "plateau-creative": { "command": "npx", "args": ["-y", "@yodolabs/plateau-creative-mcp"], "env": { "PLATEAU_AUTO_DOWNLOAD": "true", "PLATEAU_OUTPUT_DIR": "/Users/you/plateau-out" } } } }

Option B — bring your own artifacts

If you already have plateau-bridge's out_ / outputs, point straight at them:

export PLATEAU_ARTIFACT_DIR=/path/to/plateau-bridge

export PLATEAU_OUTPUT_DIR=./out npx @yodolabs/plateau-creative-mcp # stdio MCP server

{

"mcpServers": { "plateau-creative": { "command": "npx", "args": ["-y", "@yodolabs/plateau-creative-mcp"], "env": { "PLATEAU_ARTIFACT_DIR": "/path/to/plateau-bridge", "PLATEAU_OUTPUT_DIR": "/Users/you/plateau-out" } } } }

Smoke run without an MCP client:

npx tsx scripts/smoke-load-area.ts

Example workflows

1. Cyberpunk Shibuya flythrough (cross-MCP)

load_area(city="shibuya", bbox=[139.6975,35.6555,139.7045,35.6605], lod=2)

-> compose_scene(time="19:30", weather="rain") -> export_glb() # to ./out/ _v3.glb => BlenderMCP: import_glb + cyberpunk shader + render

flowchart LR

LLM(("LLM")) --> A["plateau-creative-mcp load · edit · export_glb"] LLM --> B["Blender MCP import · shade · render"] A -->|".glb + sidecar"| B

2. Tokyo without skyscrapers

load_area(...)

-> filter_buildings({height_min: 100}) -> delete_buildings(filter={height_min: 100}) -> export_glb()

3. Office-tower draft (requires zoning fields in the artifact)

load_area(...)

-> filter_buildings({zoning_use: ["commercial"], far_max_min: 400}) => SketchUp MCP: design a new building inside those parcels

Architecture

src/

server/ MCP protocol entry, tool registry, unified envelope tools/ 10 MCP tool handlers schemas/ Zod schemas + JSON Schema export scene/ scene_id LRU + lock + optional disk persistence data/ Artifact (duckdb) access layer + on-demand bundle downloader export/ three.js scene builder + GLTFExporter + manifest fallback attribution/ Envelope wrapping, attribution merge rateLimit/ Token bucket per (client, tool) errors/ Stable error codes + AppError config/ Env-driven config utils/ ids, bbox, paths, logger

• `scene_id` is the only shared state. All mutate / export tools must reference it. IDs are server-generated ULIDs.
• `building_uid` is the single canonical reference. No gml_id or batch index alternates.
• `export_glb` hard limits: ≤ 1 km² bbox and ≤ 5000 buildings for single_glb. Larger scenes return EXPORT_LIMIT_EXCEEDED with a suggestion to switch to scene_manifest.
• Attribution is wrapped at the executor level — no handler can return a result without it.

Configuration

See .env.example for a copy-paste template.

Footprint geometry vs box fallback

Polygon extrusion via the DuckDB spatial extension is on by default; on first run the extension is auto-installed. If install/load fails (offline CI, restricted environment), the exporter cleanly falls back to centroid-anchored 12 m × 12 m boxes per building. available_attributes in load_area's response includes footprint_polygon when the high-fidelity path is active.

Mesh merging + sidecar index

single_glb exports merge every building into on