RFD 1095: A voice XR path, beside Task Manager, same backend
Problem
Task Manager already drives 3DAIGC-API through REST, but a headset user with hands full, or wanting a spoken “make a 3D model of this,” has no voice-and-camera path to the same inference backend.
Decision
Wire NVIDIA’s xr-ai stack, running on the CUDA GPU host, to 3DAIGC-API through an HTTP MCP adapter, not a second inference backend. Any WebXR headset’s camera and microphone reach the XR Media Hub (port 8088, LiveKit for WebRTC on 7880-7882), which runs speech-to-text, a vision-language model, then text-to-speech; a voice-agent sample (3daigc-vlm-example) calls the same 3DAIGC-API mesh jobs through MCP tools (upload_image, image_to_textured_mesh, wait_for_job). When a router blocks the headset from reaching the GPU host directly, scripts/xr-spark-hub-proxy.mjs on the editor workstation forwards :8443 to the GPU host’s own :8088. RFD 1086’s Surface and DGX Spark are this team’s own reference pair, tested with Galaxy XR; RFD 1119 gives the general requirement.
See DETAILS.md for the full port table, the start and monitor scripts, the MCP tool flow, and the troubleshooting table.