NodeArchitect AI Agent - AI Agent and MCP Server for Houdini FX
byā¢
NodeArchitect is an AI agent and MCP server for Houdini that reads your live node graph, parameters, and connections. Ask it to explain a network, debug an error, write VEX/Python, or build an HDA ā grounded in your actual scene, not guesses. v2.3.1 adds ChatGPT/Codex login, Plan Mode, HDA Composer and Copernicus support. v2.4.0 adds Houdini 22 support.
Replies
Last time I posted here, NodeArchitect was a much smaller tool ā basically a chat window bolted onto Houdini. Since then it's had a full rework (2.0) and a run of updates since, and it's a genuinely different product now:
What changed since we last posted:
ā Houdini 22 Support
ā UI overhaul. The interface used to be built natively in Houdini's Python/Qt ā it's now a web-based UI instead, which gives much better text rendering (proper markdown, code blocks, syntax highlighting) and makes the whole thing easier to iterate on.
ā Full scene-aware agent. It reads your node graph, parameters, and connections directly, so answers are grounded in what's actually in your file instead of generic Houdini advice.
ā Built-in MCP server, no separate setup. External clients like Claude Desktop and Cursor can also connect to your live Houdini session if you want them to.
ā Per-tool consent. The agent asks before anything mutates your scene ā nothing happens silently in the background.
ā Multi-provider support, not locked to one LLM.
ā Persistent chat sessions with AI-generated titles, so you can pick a conversation back up later.
ā HDA Building ā describe a tool in plain English, get a Python build script.
And what's new in the latest release (v2.3.1):
ā ChatGPT/Codex login via the Codex CLI. You can try it with an existing ChatGPT or Copilot sub instead of setting up a separate API key ā this was the biggest friction point before.
ā Plan Mode. The agent reads your live scene and puts together a full plan before it changes anything. You review, edit, or bin it before a single node moves.
ā HDA Composer. Same idea as HDA Building, but plan-first with rollback if something breaks partway through.
ā First Copernicus support ā 23 tools for COP node inspection, image planes, diagnostics, and exporting comps. Early days for this part, more coming.
ā Project memory across sessions, cook-time profiling, undo-safe edits.
I build VFX pipelines for a living, so this is built around how I'd actually want an agent to behave inside a scene I can't afford to break ā plan first, ask before mutating, stay grounded in what's actually there instead of hallucinating node names.
Happy to answer anything ā Codex auth flow, Plan Mode internals, Copernicus roadmap, what the 2.0 rework actually involved, whatever.