Artist-controlled AI assistance inside Houdini.
Describe a task. Build editable nodes. Keep the creative decisions.

_{v5.67.4 · Houdini 22.0.400 · Python 3.13 runtime
tags: v5.67.4 is Latest}

Start here: Install · First build · What's ready · Help

The idea

SYNAPSE helps with the routine work of building and inspecting Houdini networks, so you can spend more attention on the parts that make a shot yours.

Think cruise control: you decide when help is useful and what you want to make. That is our artist-first intent. Full predictive creation and Computer Use are future capabilities; today's starting point is explicit requests and a small, optional checked-suggestion scaffold.

Start here

Current validation target: Windows with Houdini 22.0.400 and its bundled Python 3.13. Other builds need their own checks.

1. Download and extract the source ZIP from the latest release into a folder you will keep.

2. Open a terminal in that folder and run both installers:

   python scripts/install_synapse_package.py
   python install.py

3. Restart Houdini. Open New Pane Tab → Synapse.

4. Click Connect models, choose an engine and model, select Build and edit networks, then Check connection → Use this model.

For cloud engines, enter the API key in the connection dialog. For a local model, start Ollama first and choose an installed model. The connection check reads metadata; it does not test a build. Tool support varies by model.

Two different controls: Connect models chooses your AI service. Connect starts Houdini's local bridge. Doctor checks SYNAPSE without asking a model. Full setup and troubleshooting →

Your first build

Start in a scratch scene. Send:

make a box

Inspect the result in Houdini, then try Undo. Undo operates on a recorded operation; it does not reverse an entire conversation or external file writes.

For the Solaris demo, start with “create a solaris network”, then specify your geometry, HDRI, lights, MaterialX material, camera and Karma XPU settings. Check the resulting nodes and render configuration before treating it as shot-ready.

First-session walkthrough →

flowchart TD
    A["You describe a task"] --> B["SYNAPSE uses allowed tools"]
    B --> C["You inspect and edit the nodes"]
    C --> D["Keep, refine, or undo an operation"]

Loading

Recorded demonstration; the current panel and feature limits are described below.

What's ready

Feature	Current state
Build and inspect	Explicit requests through the panel's permitted tools; editable Houdini nodes.
Choose your model	Claude, Gemini, NVIDIA Nemotron, Ollama and a custom OpenAI-compatible endpoint. Capabilities vary by model.
Doctor and panel controls	Direct diagnostics, visible Connect / Doctor controls, and installed Ollama model discovery.
Saved lookdev suggestion	Optional Stage 0 scaffold for one checked workflow. It prepares an editable prompt; you decide whether to send it.
Memory LOOP	Optional observation and recall across Moneta, Octavius and Hanish. Requires configured substrates.
Predictive creation + Computer Use	Product direction in INTENT.md; independent artist controls are planned.
Recursive self-improvement	A development direction. Stage 0 is a starting scaffold, not a self-maintaining system.

How the three memory substrates work together

flowchart TD
    M["Moneta: remember project outcomes"] --> O["Octavius: compose private context"]
    O --> H["Hanish: record forecast and outcome"]
    H --> M

Loading

SYNAPSE coordinates this optional loop around requested operations. Recalled records remain advice; they do not choose the next scene action. Unavailable substrates and unknown outcomes remain visible.

Saved suggestions require an initialized project store and a compatible, developer-imported record. Use lightning tools menu → Saved lookdev suggestion… or /lookdev-suggestion. Use in prompt prepares text without sending it. Exact version matching means an older checked experience needs a new rehearsal and import after a SYNAPSE version change.

LOOP setup · Stage 0 setup · Detailed diagrams

When you get stuck

What you see	First thing to try
No Synapse pane	Restart Houdini, then follow installation verification.
Missing Ollama models	Start Ollama at the configured address, reopen Connect models, then Check connection.
Houdini connection trouble	Click Connect, then Doctor.
No saved suggestion	Read the card's reason. A matching imported experience and an available project memory owner are required.

Know the limits: a failed build can leave partial nodes; undo grouping does not guarantee rollback. Stop prevents further panel work, but does not prove that a running cook or render has finished. The external bridge has broader permissions than normal panel chat and is intended for a single-user local machine.

Current limits and verification · Report a bug

For developers: architecture, evidence and contribution

128 tools, two paths. The registry is the authority. HTTP MCP and WebSocket execution have different policy boundaries; the panel's worker adds its own tool restrictions and model-data permissions.

• Architecture and Mermaid diagrams
• Developer installation and tests
• MCP client setup
• Changelog and release notes
• Contribution rules and execution contracts

GitHub CI runs the stock-Python suite on Linux and macOS. Native Houdini checks are a separate qualification; a green CI badge is not a live render or demo pass. Release tags are created after version agreement and published after their exact commit passes CI. Release process →

License

MIT. Patent applications pending on the USD cognitive-state substrate, digital injection, and predictive lighting.