I'm a solo developer and VFX artist. Houdini is the tool I use most, and for the few months I've been trying to make AI actually useful inside it.
The options that existed were either priced for studios or required too much friction to get running. Houdini-mcp, for example, needed code changes just to work on the latest version of Houdini. So I built NodeArchitect.
You select nodes in your scene, ask a question, and get an answer grounded in your actual node graph — not a guess based on a generic prompt.
The things it's most useful for:
"What is this network doing?" — useful when you inherit someone else's scene
"Why is this breaking?" — debugging with full parameter context
"Generate VEX or Python for this" — with your specific node setup in view
"Turn this into a reusable HDA"
"Document this for handoff"
The original version used text-parsing to interpret what the agent should do. It worked, but it was fragile — multi-step tasks would lose context, and debugging was unreliable.
2.0 is a full rewrite using native function calling. The agent now routes tool calls properly, and multi-step tasks are significantly more reliable as a result.
Other changes in 2.0:
New web UI — replaced the old PySide panel with a proper embedded web frontend. Streaming responses, syntax-highlighted code, HiDPI support.
Built-in MCP server — NodeArchitect runs on its own bundled MCP server, no separate install. External clients like Claude Code can also connect to your live Houdini session.
GitHub Copilot login — sign in with an existing Copilot subscription (including the free tier). No API key, no metered billing.
Per-tool consent — the agent pauses before anything mutates your scene. You can approve, edit, or reject before it runs.
Multimodal — you can now upload images (concept art, references) as context. The MCP server can also snapshot the Houdini viewport mid-task so the agent can see what it's doing.
Download and more info: https://bishopvfx.gumroad.com/l/houdininodearchitect