houdini MCP server - dev v9 progress tests

These are a couple of live sessions I just did with the development version 9 of my MCP server. The progress has been both swift and substantial. The LLM's capabilities to work inside Houdini is improving noticeably version for version. Although there is much more under the hood, here are some current highlights for what it can already do: **1. Build node networks by talking** Claude can create nodes, wire them together, set parameters, and lay them out in the network editor — all from a chat conversation. You describe what you want, it builds the graph. Works for SOPs, COPs (Copernicus), and object-level networks. **2. Run Python and VEX inside Houdini** Claude can write and execute Python scripts and VEX snippets directly in the running Houdini session — with built-in safety checks so it doesn't do anything destructive by accident. **3. Copernicus (COP) image workflows** Full support for Houdini's new Copernicus compositing: create networks, chain nodes together, read and write pixel data, inspect layer metadata. Claude can build COP graphs including pyro simulations. **4. Viewport rendering and visual feedback** Claude can capture the viewport as an image, look at it, and adjust — a render-check-adjust loop. It can also trigger ROP renders, track their progress, and cancel them. Since Claude can actually see the renders, it can iterate on framing and look-dev. **5. Live scene awareness** The server provides real-time feedback: geometry stats (point/prim counts, attribute statistics), error diagnostics across the full node tree, and an event system that notifies Claude when nodes cook, get renamed, or have parameter changes. **6. Long-running operation management** Heavy operations (renders, large cooks) run in the background. Claude can poll progress and cancel if needed — it doesn't block the Houdini session while waiting. **7. Progressive tool discovery** 54 tools are available, but Claude only sees 5–8 at a time based on what it's doing. This keeps the AI focused and prevents it from being overwhelmed by options — important for getting reliable, predictable behavior from the LLM. #houdini #vfx #ai #agentic #agenticcoding #agenticai #mcp #mcpserver #visualeffects #3d #proceduralism