An MCP server that lets AI assistants read, inspect, and manipulate [OpenUSD](https://openusd.org) scenes.
openusd-mcp
An MCP server that lets AI assistants read, inspect, and manipulate OpenUSD scenes.
Live Demo — interactive 3D viewer with variant switching
pip install openusd-mcp
What it does
Connect this MCP server to Claude, ChatGPT, Cursor, or any MCP-compatible client. Your AI assistant can then work directly with .usda, .usdc, and .usdz files — inspecting scene graphs, querying properties, exporting meshes, and switching variants.
Tools
| Tool | Description |
|------|-------------|
| usd_inspect | Read the scene graph — list all prims, types, hierarchy |
| usd_get_prim | Get detailed properties, attributes, and metadata for a specific prim |
| usd_get_materials | List all materials and their shader parameters |
| usd_get_transforms | Get world-space transforms for prims |
| usd_list_variants | List all variant sets and their options |
| usd_set_variant | Switch a variant selection |
| usd_export_mesh | Export a mesh prim as STL or OBJ |
| usd_scene_stats | Get scene statistics — prim count, mesh count, material count, bounding box |
Quick start
With Claude Desktop
Add to your claude_desktop_config.json:
{
"mcpServers": {
"openusd": {
"command": "openusd-mcp",
"args": []
}
}
}
Then try it with the included example scenes:
"Inspect the scene graph of examples/desk_setup.usda and describe what's on the desk."
"What color options does the product in examples/product_configurator.usda have? Switch it to gold."
With any MCP client
openusd-mcp --stdio
The server communicates over stdin/stdout using the MCP protocol.
Examples
The repo includes two example scenes that exercise all 8 tools.
Desk setup — scene inspection
A desk with a table, monitor, keyboard, and mug. 5 PBR materials (wood, metal, screen, ceramic, dark plastic), 10 meshes, real-world dimensions.
You: What's in examples/desk_setup.usda?
Claude: A desk scene with 25 prims, 10 meshes, 5 materials:
/Desk (Xform)
/Desk/Table/Top (Mesh) — Wood, 120×60×3 cm
/Desk/Table/Leg1..4 (Mesh) — Metal
/Desk/Monitor/Screen (Mesh) — dark glossy
/Desk/Monitor/Stand (Mesh) — Metal
/Desk/Monitor/Base (Mesh) — Metal
/Desk/Keyboard (Mesh) — DarkPlastic
/Desk/Mug/Body (Mesh) — Ceramic
Bounds: 1200 × 835 × 1250 mm
You: Export the tabletop as STL
Claude: Exported /Desk/Table/Top to tabletop.stl (12 triangles, 684 bytes)
Product configurator — variant switching
A device with 3 variant sets: size (small, medium, large), color (midnight, silver, gold, red), and finish (matte, satin, glossy).
You: What configuration options does examples/product_configurator.usda have?
Claude: The product at /Product has 3 variant sets:
size: small | medium* | large
color: midnight* | silver | gold | red
finish: matte | satin* | glossy
(* = currently selected)
You: Switch to gold, large, glossy
Claude: Done. Updated /Product:
size: large (1.4× scale)
color: gold — diffuseColor: (0.83, 0.69, 0.22), metallic: 0.9
finish: glossy — roughness: 0.1
Requirements
- Python 3.9+
usd-core(Pixar's USD Python bindings)
pip install usd-core
How it works
The server uses Pixar's official pxr Python bindings (the same libraries used by NVIDIA Omniverse, Apple's Reality Composer, and every major VFX pipeline) to read and manipulate USD stages. It exposes these capabilities as MCP tools that any AI assistant can call.
OpenUSD is the industry standard for describing 3D scenes, backed by Pixar, Apple, NVIDIA, Adobe, and Autodesk. As of December 2025, the Alliance for OpenUSD has ratified Core Specification 1.0.
Author
License
MIT