What I was connecting
MCP was the connection through which an assistant could operate my 3D applications. The useful capability varied with the bridge: reading a scene, executing Python, constructing a node network or exposing editor operations.
I tested several approaches over the course of these projects. A connection being installed or responding was only the first check; the next question was whether it exposed the operation the task actually needed.
The difference I noticed
Limited commands for spawning and transforming objects were often too coarse for the work I wanted. Access to lower-level operations let the assistant construct a reusable tool, inspect its data and revise it in smaller steps.
That flexibility also required more checking. Long operations timed out, application versions changed what was available, and correct data in one application could be stale in the next.
How this appears in the projects
- The ruined wall used graph construction to turn five authored stones into an editable prototype.
- Organic architecture connected a Houdini system to controls I could use in Unreal.
- Market Street tested the harder problem of maintaining a complete transfer across applications.
My notes separate a working connection, a working tool and a repeatable pipeline. Each needs its own evidence.