Joseph Bishop

Hi everyone,

NodeArchitect 2.4.0 is now available.

This update is focused on Houdini 22 support, but it is not a hard Houdini 22-only release. NodeArchitect now detects the live Houdini build and available nodes before exposing specialised workflows. Houdini 21 continues to use compatible tools and automatic fallbacks where newer Houdini 22 functionality is unavailable, including Karma AOV output behaviour.

New in 2.4.0:

• APEX and KineFX workflows for rig inspection, skeleton debugging, retargeting and character recipes.

• Houdini 22 Copernicus workflows for terrain, pyro and material recipes.

• Solaris and Karma additions: RenderPasses, AOVs, Texture Material Library support, preview-shot setup and viewport-to-camera capture.

• PDG, ML and Gaussian Splat workflows for dataset inspection, training setup and diagnostics.

• SOP camera and implicit-surface workflows.

• Context-sensitive MCP tool discovery. The built-in agent and external clients now receive tools relevant to the active SOP, LOP or TOP context, keeping the tool list focused rather than bloated.

v2.4.0 has been smoke-tested against Houdini 22.0.368 and Houdini 21.0.700.

Download / update [bishopvfx.gumroad.com]

Big update: NodeArchitect has now been updated to v2.3.x


ChatGPT/Codex login is now supported via the Codex CLI. This should make NodeArchitect much easier to try if you already use ChatGPT or Codex.

This release also adds Plan Mode, which is a major change to how the agent works. Before making any changes to your Houdini scene, NodeArchitect now generates a full implementation plan using live scene data. You can review the plan, refine it, or discard it entirely. Nothing executes until you approve it.

There is also a new HDA Composer. The agent can now help create, modify, and validate HDAs from natural language prompts. Like Plan Mode, it works plan-first, with rollback handling if something goes wrong.

Another big addition is early Copernicus / COP support. NodeArchitect previously very little Copernicus integration, so this release increases the COP toolset: 23 tools covering node inspection, image plane queries, error diagnostics, and exporting comps to disk.

Other improvements in v2.3.1 include:

• Project memory - NodeArchitect now retains context scoped to the project file rather than just the user session. Returning to a scene, picks up prior context instead of starting cold.
  
• Cook-time performance profiling - new tool profiles the network and ranks nodes by cook time, surfacing bottlenecks directly without manually digging through Performance Monitor.
  
• Undo-safe AI mutations — every AI-driven scene edit now creates a checkpoint automatically, so changes can be rolled back from the NodeArchitect panel.
  
• Compact conversations — long sessions are automatically summarized before hitting the model’s context limit, while preserving important decisions and exact Houdini paths.
  
• Better model compatibility — text-only models now fall back to text-based validation instead of breaking on viewport requests.
  
• Stability fixes — including cook-deadlock prevention for agent-triggered scene evaluation and other reliability improvements.
  

Undo Changes Safely

Overall, this is one of the biggest NodeArchitect updates so far. The focus has been on making the agent safer, easier to get started with, and more useful in real Houdini workflows: ChatGPT/Codex login, plan-first execution, rollback-safe edits, HDA creation, longer sessions, and the first step toward COP/compositing support.

I previously built NodeArchitect. While developing it, I found that giving an assistant access to Houdini only solves part of the workflow.

It may understand the current Houdini session, but not which asset or shot it belongs to, which input version is current, where the .hip file should be saved, or what downstream work depends on it. That context can also disappear between sessions.
I built KeystoneMCP to provide that missing production layer. It is a local MCP server aimed primarily at solo artists and small creative projects.

KeystoneMCP lets an AI assistant manage:
Project folder creation
Tasks, assets and shots
File locations and naming rules
Work, review, publication and delivery files
Current and approved versions
Dependencies and blocked tasks
Project context across sessions



It works alongside Houdini MCP tools rather than replacing them. The Houdini MCP operates inside Houdini, while KeystoneMCP provides the wider project state, exact paths, input versions and workflow guidance.

The goal is to let AI handle more of the repetitive project organization while the artist remains focused on the creative work.

KeystoneMCP is currently free to claim for eligible indie users and businesses earning under $150,000 USD annually. Anyone eligible who claims it during the launch period keeps that version permanently. A low-cost Indie tier will follow in roughly one month.

Download [bishopvfx.gumroad.com]