HoudiniMCP

Houdini is a vast piece of software, with oddities everywhere waiting to be found.

It's insanely powerful, but sometimes, it can get a bit overwhelming.

Asking LLMs for advice can work, but frontier models still struggle to come up with proper VEX syntax, and rely too much on obscure un-editable wrangles to create procedural setups.

The HoudiniMCP gives agents a way to step inside your scene and tweak your nodes, take screenshots, and iterate until the setup is up to your standards, both artistically and technically.

Many MCPs exist for Houdini. This one stands out by being slightly opinionated (prefers built-in nodes rather than wrangles, avoid single-letter variable names, …), having a simplified tool schema for reduced token usage, and strong visual verification tools (playblasts, references, …)

I recommend setting up a "Stop" hook in your harness of choice, since agents are bad at objectively judging subjective works like visuals, and are quick to call something "done". Setting up a hook would re-prompt them as soon as they stop, until they cannot see anything to improve anymore. At that point, the same hook could also prompt them to ask another subagent for feedback, and keep iterating. I have seen agents go on 16+ hrs iterative runs on scenes, and the results were mindblowing.

Frontier models are still not better than humans for visual taste. What they are good at is iterating when you, the human, is not available or are focusing on something else.

Houdini is a vast piece of software, with oddities everywhere waiting to be found.

It's insanely powerful, but sometimes, it can get a bit overwhelming.

Asking LLMs for advice can work, but frontier models still struggle to come up with proper VEX syntax, and rely too much on obscure un-editable wrangles to create procedural setups.

The HoudiniMCP gives agents a way to step inside your scene and tweak your nodes, take screenshots, and iterate until the setup is up to your standards, both artistically and technically.

Frontier models are still not better than humans for visual taste. What they are good at is iterating when you, the human, is not available or are focusing on something else.