Inspect CAD images in LM Studio
Set up the optional vision adapter and distinguish actual image feedback from file references.
Updated:
Vision capability and image delivery
A direct local AI chat in ShapeAnywhere sends actual CAD images to the local LM Studio server without this adapter. The instructions below remain an optional solution for LM Studio's own chat. A vision model can inspect images when the client supplies actual image inputs. A filename or visible chat preview does not establish this. Some LM Studio versions save MCP images while forwarding only a file reference to the model. Assess actual visual descriptions and precise measurements separately.
Install the optional adapter
The local ShapeAnywhere Vision adapter restores this input through LM Studio's normal plugin interface. LM Studio and its CLI must already be installed. No additional ShapeAnywhere access token is needed. Download ShapeAnywhere Vision for LM Studio 1. Extract the entire ZIP. On Windows run Installieren.cmd; alternatively run lms dev --install from the extracted directory. Installation may download npm dependencies, but no language model. 2. Select shapeanywhere/vision under Your Generators in the model menu. This is the adapter; the underlying model remains local. 3. Keep the existing mcp/shapeanywhere integration enabled. Open the intended CAD project with at least one CAD solid and connect MCP. 4. Enter the local model identifier in generator settings. The default is qwen/qwen3.5-4b. Other locally available vision models can be selected/entered there. Language, image limit, temperature and response limit are configurable.
Read-only first test
Call cad_get_view exactly once with view="isometric", width=768 and height=576. Then describe only visible shapes and colors. Use no other tools and do not change geometry.
A concrete observation might identify a cylinder with a rim and visible holes. Dimensions and hidden features need additional checks. No model image can be produced without an open CAD solid. Client and CAD approvals still apply.
Behavior and limits
Before the next response, the model receives up to three recent CAD views as actual image inputs; one to four are configurable. Images remain chronologically after their tool replies so later model edits stay distinguishable. The visible tool text may still contain a filename. Only recognized PNG previews from the current chat's working directory are read. Missing images need a fresh request. The adapter executes no tools itself and changes neither solids nor approvals. Select the original model at any time to return to ordinary inference. Successful vision does not replace geometry, mechanism or print checks. The native flow was tested on Windows with LM Studio 0.4.28 and Qwen3.5-4B. Continue with MCP connection and tools and checking free AI construction.