Alexander Rybashov

Hi everyone,

I’ve been testing Houdini MCP with Codex Desktop on Windows and put together a working setup for controlling a live Houdini session from an AI coding agent.

The setup lets Codex connect to Houdini through MCP and hrpyc, so the agent can inspect and edit the current scene: create nodes, wire networks, set parameters, run Python in the open Houdini session, and work with Solaris/USD graphs.

I also opened a pull request with a couple of fixes that were needed to make this reliable on Windows:

- keep FastMCP stdio output clean by disabling the banner
- run stdio without HTTP host/port arguments
- start Houdini’s hrpyc listener through ThreadedServer
- stop the hrpyc server via the server object instead of the missing hrpyc.stop_server API
- document a tested Codex Desktop + Houdini 21.0 Windows setup

Project page:
https://arybashov.github.io/houdini-mcp/ [arybashov.github.io]

Fork:
https://github.com/arybashov/houdini-mcp [github.com]

Upstream PR:
https://github.com/oculairmedia/houdini-mcp/pull/8 [github.com]

Telegram updates:
https://t.me/cg_pipeline_lab [t.me]

Tested with:
- Windows 11
- Houdini 21.0.671
- Codex Desktop
- Python 3.14
- RPyC 5.3.1

The most stable architecture so far is:

Codex Desktop
-> stdio MCP server running in normal Python
-> RPyC 5.x
-> Houdini hrpyc listener on 127.0.0.1:18811

I’m sharing this in case other Houdini users are experimenting with AI agents, MCP, Solaris/USD automation, or DCC pipeline tools. Feedback and testing on other Houdini versions would be very useful.