AI Assistant¶
The AI Assistant is a panel in usdview for chatting with Claude, OpenAI,
Grok or a local model about the open stage and its rigs. It looks at the
stage with tools rather than guessing, and with your permission it poses and
keys controls, edits the stage and runs Python. Each step it takes is listed
in the chat as it works.
It is off until you turn it on. RigExec ▸ General Editors ▸ AI Assistant opens its settings the first time; check Turn on the AI assistant and the panel docks on the right of the window. Nothing is sent to a provider until you ask it something.
The panel¶
- Provider and model. Who answers, and which of their models. The line under them says which sign in or key is used. Settings opens the settings.
- Chats. Every chat is kept, so New Chat does not lose the last one. Click a chat to go back to it. Right-click a chat to Pin it to the top or Delete it. Chats are saved on this computer only.
- The conversation. Your messages and the replies. Each reply lists the steps it took, with a dot for each kind: blue for looking, purple for selecting prims or changing the frame, green for changing the stage, orange for running Python, and a red cross for a step that failed. What it changed is listed in a green box under the reply.
- Run Code and Copy Code appear when a reply has a block of Python. Run Code runs it in usdview and needs Run code on.
- The message box. Ctrl+Enter sends. While a reply is coming, Send becomes Stop.
- What it may do. These toggles are remembered:
- Selection sends the selected prims and their attributes with your message.
- Edit stage lets it pose and key controls, make, change and delete prims, attributes, connections, relationships, references and layers, and save layers. Edits go to the stage's edit target.
- Run code lets it run Python in usdview to test its work and read what it prints. It is off until you turn it on, and shows orange when on.
All the changes in one reply are one step on the RigExec undo stack, so Ctrl+Z takes the whole reply back. Python it runs can be undone too, as long as it only changed the edit target.
Signing in¶
The settings have a section for each provider. Click a section's header to open it; the header says how that provider is set up.
- Claude uses the Claude Code app's sign in, so any Claude plan works and usdRig never sees the account's token. Sign in with Claude Code runs Claude Code's own sign in in the browser. Claude Code needs to be installed.
- OpenAI can sign in with a ChatGPT plan, Codex uses the Codex app's sign in, and Grok can sign in with SuperGrok.
- Each can use an API key instead, or the usual environment variable
(
ANTHROPIC_API_KEY,OPENAI_API_KEY,XAI_API_KEY). - Local Model points at a server on this computer or the network: Ollama, LM Studio, llama.cpp, vLLM or Jan. Find looks for one that is running.
Ask ... first picks the provider the panel starts with, and Check asks a provider for its models. Keys and sign ins are kept in the operating system's secure store.
What it can do¶
| Kind | Tools |
|---|---|
| Look at the stage | The stage overview (layers, edit target, time range, rigs), prim outlines, a prim or attribute in full (values, keys, connections, the layers with opinions), searches by name, type, schema or attribute. |
| Look at a rig | The rigs on the stage, a control's avars, compiling a rig and reporting its errors and mover order, evaluating it at a frame for joint and control frames, and what the viewport's own evaluator last did. |
| Look at the viewport | A picture of the viewport, for the providers that can see images. |
| Show you | Select prims (and frame them), go to a frame. |
| Pose | Set and key avars, or reset them to the rig's rest. |
| Change the stage | Define, remove, rename and move prims; set, key, clear and block attributes; connections, relationships, schemas, variants, references, sublayers, the edit target; save layers. |
| Run Python | Python in usdview, with usdviewApi, stage and rigexec ready. |
| Read files | Files in the stage's folder, and anywhere else you allow. |
check_rig and evaluate_rig compile the rig on their own, apart from the
viewport, and author nothing. Avars that a mover recomputes are reported as
hidden: setting them changes nothing you can see, so the assistant poses the
controls that drive them instead.
With a Claude or Codex sign in, the tools are served to Claude Code or Codex by the assistant's own MCP server. It only listens on this computer and takes a new key for every reply. The other providers use the same tools through their own tool calling. A model that cannot use tools answers without them, and changes the stage with a block of changes the panel applies.
Reading files¶
The assistant can read files on this computer. The open stage's folder is free to read. Before it reads anywhere else it asks, with its reason:
- Allow lets it read that folder for the rest of the chat.
- Always allow lets it read the folder in every chat, and remembers it under Memory in the settings.
- Deny says no, and it does not ask again in that reply.
Memory¶
Ask the assistant to remember something, such as how you name controls, and
it saves a short note. Notes are sent at the start of every chat. Memory
in the settings lists them and the folders it can always read. Settings,
chats and notes are kept in ~/.usdRig/assistant; set USDRIG_USER_DIR to
keep them elsewhere.