Houdini-Agent is an AI-powered assistant for SideFX Houdini, designed to automate complex DCC asset management tasks. It targets Houdini artists and technical directors, offering autonomous multi-turn tool calling, code execution, and web search to streamline workflows and enhance productivity. The agent leverages OpenAI Function Calling and a centralized ToolRegistry to interact with Houdini's node networks, scripts, and external resources.

How It Works

The core of Houdini-Agent operates in an autonomous agent loop: receiving a user request, planning steps, calling tools, inspecting results, and iterating until completion. It supports three distinct modes: Agent mode for full operational control, Ask mode for read-only queries, and Plan mode for structured, user-reviewed execution plans with DAG visualization. Its approach integrates a brain-inspired long-term memory system, a plugin hook system for extensions, and user-defined context rules, all managed via a unified ToolRegistry.

Quick Start & Requirements

• Installation: Clone or download the repository. No pip install is required as dependencies are bundled. Add the repository's root directory to sys.path in Houdini and run import launcher; launcher.show_tool(), or add this to a Shelf Tool.
• Prerequisites: Houdini 20.5+ (or 21+), Python 3.9+ (bundled with Houdini). PySide2 (Houdini ≤20.5) or PySide6 (Houdini 21+) is required. Windows and macOS are tested; Linux support is possible.
• Configuration: API keys can be configured via environment variables or the in-app settings dialog.
• Links: Plugin development documentation

Highlighted Details

• Plan Mode: Generates structured execution plans with phases, steps, dependencies, risk assessment, and DAG flow diagrams for user review and monitoring.
• Long-Term Memory: A three-layer (episodic, semantic, procedural) memory system with reward-driven learning and automatic reflection enables continuous improvement.
• Plugin System: Supports community extensions via a plugin architecture with event hooks, custom AI-callable tools, UI buttons, and settings management.
• Vision/Image Input: Supports multimodal messages with image attachments (PNG/JPG/GIF/WebP) for vision-capable models.
• Token Analytics: Real-time tracking of token counts, reasoning tokens, cache hit rate, and per-model cost estimates.
• Advanced Tools: Includes capture_viewport, analyze_cook_performance, create_wrangle_node with VEX, execute_shell, and extensive node inspection/manipulation tools.

Maintenance & Community

The README does not detail specific contributors, sponsorships, or community channels like Discord/Slack. Updates are managed via an auto-updater checking GitHub Releases.

Licensing & Compatibility

The project is licensed under the MIT License, which is permissive for commercial use and closed-source linking.

Limitations & Caveats

While tested on Windows and macOS, Linux support is noted as "possible" but not explicitly tested. The agent's effectiveness is dependent on the underlying AI model's capabilities and the accuracy of configured API keys.