Claude × Houdini MCP Integration | How to Create VFX Scenes with Natural Language

What is HoudiniMCP?

HoudiniMCP is a third-party bridge that connects Anthropic's Claude (Desktop version) to SideFX's Houdini via the Model Context Protocol. Developer capoomgit has released it as open source on GitHub (capoomgit/houdini-mcp), and it is one of a growing number of MCP-compatible tools that have surged since 2024.

The system uses a two-layer architecture. A socket server (Python script) running inside Houdini listens on port 9876, receives commands from a bridge process launched by Claude Desktop, and calls Houdini's Python API. From Claude's perspective, Houdini functions as a "scriptable execution environment you can talk to," enabling 168 types of operations to be issued with text instructions alone — including node creation, geometry generation, simulation control, and rendering (according to mcpworld.com's explanation).

MCP itself is an open protocol defined by Anthropic, serving as the standard specification for connecting external tools to Claude. By building on this specification, HoudiniMCP is designed to maintain compatibility even as Claude's capabilities are enhanced in the future.

Why Houdini is Difficult — and How Claude Changes That

Houdini is a professional-grade tool used by VFX studios in film, commercials, and game effects production, and it has an extremely steep learning curve. Its node-based procedural workflow is powerful, but the sheer number of settings for particle, fluid, and destruction simulations is overwhelming — it is not unusual for beginners to spend hours just building a simple scene of smoke rising from the ground.

Furthermore, while Houdini's Python API is extensive, just researching the documentation to connect the right nodes in the right order is a challenge in itself. Connecting Claude dramatically lowers this barrier. In response to text instructions like "Set up a scene where an explosion goes off 10 meters away and debris scatters," Claude automatically generates the necessary nodes and configures their parameters. The ability to quickly prototype scenes without deep Houdini expertise is a major advantage for both professionals and amateurs alike.

Prerequisites

Confirm that all four of the following are in place:

1. SideFX Houdini 19.0 or later (19.5 recommended)
2. Python 3.10 or later (installed on your system)
3. uv package manager (used to manage MCP dependency libraries)
4. Claude Desktop (latest version)

Houdini Apprentice (the free version) is technically said to work as well, but note that Houdini's own restrictions — such as no commercial use and limited FBX export — remain in effect with Apprentice. A paid license is required for serious professional use.

Installation Steps

Installing uv

On Windows, run the following in PowerShell:

powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

After installation, add uv to your PATH, then add the MCP library with the following command:

uv add "mcp[cli]"

For Mac / Linux, follow the steps in the official uv documentation (https://docs.astral.sh/uv/).

Placing the Houdini Plugin Files

Clone the capoomgit/houdini-mcp repository and place the houdinimcp folder at the following path:

• Windows: C:\Users\<YourUsername>\Documents\houdini19.5\scripts\python\houdinimcp\
• Mac: /Users/<YourUsername>/Library/Preferences/houdini/19.x/scripts/python/houdinimcp\
• Linux: ~/houdini19.x/scripts/python/houdinimcp/

The files to place are: __init__.py, server.py, houdini_mcp_server.py, and pyproject.toml.

Starting the Server Inside Houdini

1. Launch Houdini and open the Python Shell (Script Editor).
2. Run the initialization code to start the MCP server on port 9876.
3. Optionally, create a "Toggle MCP Server" shelf tool so you can start and stop it with a single click in the future.
4. Placing houdinimcp.json in Houdini's packages folder configures it to load automatically when Houdini starts.

Configuring Claude Desktop

Add the following block to the mcpServers section of Claude Desktop's configuration file claude_desktop_config.json:

{
  "mcpServers": {
    "houdini": {
      "command": "uv",
      "args": [
        "run",
        "python",
        "C:/Users/<YourUsername>/Documents/houdini19.5/scripts/python/houdinimcp/houdini_mcp_server.py"
      ]
    }
  }
}

Adjust the path to match your own environment. After restarting Claude Desktop with this configuration, a hammer icon will appear in the left sidebar, indicating that Houdini tools are available.

What You Can Actually Do

Here is a summary of the main operations you can request from Claude after connecting:

• Create and edit nodes — Responds to instructions such as "Create a geometry node with 5 spheres placed randomly"
• Apply materials — Set surface properties like "Apply a metal texture to the object"
• Set up cameras — "Place a camera with a view of the entire scene"
• Configure physics simulations — Initial setup for particles, destruction (Voronoi Fracture), smoke, and fire
• Execute rendering — Karma / Mantra renders can be triggered by Claude's instructions
• FBX export — Also supports export to game engines
• Direct Python execution — Run any Houdini Python script with the execute_houdini_code tool

According to skywork.ai's explanatory article, there are many reports of even inexperienced users successfully building complex procedural scenes using only text instructions, and the tool is said to be especially effective for prototyping purposes.

Example Instructions

Here are some scene-specific examples of what text to send to Claude in practice.

Explosion / Destruction: "Set up a scene using Voronoi Fracture where one corner of a building explodes and the outer wall collapses. Apply rigid body dynamics to the debris."

Smoke / Fire: "Set up a particle simulation of thick black smoke rising from a chimney. Make it drift to the right under the influence of wind."

Procedural Modeling: "Create a flower geometry with 12 petals radiating outward from the center. Add a subtle random tilt to each petal."

In all cases, Claude constructs the node network through Houdini's Python API, meaning you can see results without any prior knowledge of Houdini's controls — a major advantage.

Differences from Blender MCP

As a means of connecting Claude to a 3D tool, blender-mcp for Blender (ahujasid/blender-mcp) is also widely used. Here is a comparison of the main differences between the two:

Houdini is the right starting point when aiming for professional-quality VFX in film, commercials, or games. Blender is more appropriate if you want to casually explore 3D modeling or animation. Choosing based on your use case is the practical approach.

Limitations and Caveats

Here is a summary of the constraints to be aware of before use:

1. Unofficial tool — This is not officially supported by either SideFX or Anthropic. Behavior may change with software updates.
2. Only one instance at a time — The HoudiniMCP server can only run one session simultaneously.
3. Risk of arbitrary code execution — Since execute_houdini_code executes Python code directly, review your instructions carefully before sending them to avoid unintended destructive operations.
4. Break complex operations into steps — Requesting too many operations at once reduces accuracy; it is more reliable to build scenes incrementally.
5. OPUS asset integration requires a separate paid plan — The asset library integration feature requires a RapidAPI subscription.
6. Apprentice version restrictions — If using the free Houdini Apprentice, Houdini's own restrictions on FBX export and commercial use still apply.

Summary

With HoudiniMCP, you can build VFX scene prototypes through conversation with Claude alone, even without deep Houdini expertise. All you need are the four components — Houdini 19.0 or later, Python 3.10 or later, uv, and Claude Desktop — and setup takes less than 30 minutes. With 168 tools available, the ability to control everything from full-scale simulations like explosions, smoke, and destruction to procedural modeling using natural language will be a powerful asset for 3D creators.

If you'd like to try a similar approach with Blender, check out our article "Claude × Blender MCP Guide | How to Create 3D Models with Natural Language" on this site.