RFD 1095 details: architecture, ports, start scripts, and troubleshooting

Architecture

[Galaxy XR Chrome]  camera + mic
        |  HTTPS (often through a Surface proxy, see below)
        v
[xr-ai XR Media Hub :8088]  web client + LiveKit (7880-7882)
        |  speech-to-text -> vision-language model (local or NVIDIA NIM) -> text-to-speech
        v
[3daigc-vlm-example worker]  voice agent + scene vision
        |  HTTP MCP tools (upload_image, image_to_textured_mesh, wait_for_job, ...)
        v
[3daigc-mcp-http :8260]  a FastMCP adapter
        v
[3DAIGC-API :7842]  TRELLIS / mesh jobs, on the Spark GPU
        v
[this project]  loads the completed GLB/VRM in the viewport (a separate client path)

This path is complementary to Task Manager’s own UI: the same inference backend, a different, XR-native voice interface.

Repositories and paths, DGX only

Path	Role
`/home/sifr/xr-ai`	NVIDIA’s `xr-ai`: the hub, STT/TTS/VLM servers, agent samples
`/home/sifr/xr-ai/agent-samples/3daigc-vlm-example`	The voice VLM plus 3DAIGC mesh orchestrator
`/home/sifr/3DAIGC-API`	The inference API, port 7842
`/home/sifr/3DAIGC-API/mcp`	`3daigc-mcp-http` (port 8260), MCP tools over the completed API
`/home/sifr/Weftspun3DStudio/scripts/xr-spark-hub-proxy.mjs`	The optional Surface proxy, Galaxy XR to the Spark hub

Overlay config, a reference to copy: 3DAIGC-API/mcp/yaml/xr_ai_3daigc_overlay.yaml.

Ports, DGX

Port	Service
7842	`3DAIGC-API`
8260	`3daigc-mcp-http` (`/mcp`)
8088	The `xr-ai` XR Media Hub web UI, over HTTPS
7880, 7882	LiveKit (WebRTC) for the `xr-ai` hub
8443	The Surface’s own `xr-spark-hub-proxy` (optional; forwards to `https://10.0.0.158:8088`)

Starting the stack, DGX

One command starts both the DGX hub and the Surface proxy, for Galaxy XR:

bash /home/sifr/3DAIGC-API/mcp/scripts/start_xr_voice_full.sh
bash /home/sifr/3DAIGC-API/mcp/scripts/verify_xr_voice_stack.sh

DGX hub only, no Surface proxy (the headset gets a connection refused on 10.0.0.32:8443):

bash /home/sifr/3DAIGC-API/mcp/scripts/run_xr_ai_3daigc_stack.sh

Galaxy XR URL: https://10.0.0.32:8443 (the Surface proxy, forwarding to the DGX’s own :8088), never the bare LAN IP, never plain HTTP.

Prerequisites only, API plus MCP, no voice stack:

bash /home/sifr/3DAIGC-API/mcp/scripts/start_prerequisites.sh

MCP HTTP only:

bash /home/sifr/3DAIGC-API/mcp/scripts/run_http.sh

Monitor logs:

bash /home/sifr/3DAIGC-API/mcp/scripts/monitor_xr_ai_3daigc_stack.sh

Open the hub UI directly on the DGX LAN: https://10.0.0.158:8088 (accept the self-signed certificate), then start the microphone and try “make a 3D model of this.”

Galaxy XR and router client isolation

Some routers block headset-to-DGX traffic (10.0.0.224 to 10.0.0.158) while still allowing headset-to-Surface traffic (10.0.0.32).

On the Surface, using certs/localhost.pem from npm run setup-https (RFD 1088):

cd C:\Users\alfao\Documents\GitHub\Weftspun3DStudio
$env:XR_SPARK_HUB_URL = 'https://10.0.0.158:8088'
$env:XR_PROXY_PORT = '8443'
node scripts/xr-spark-hub-proxy.mjs

On Galaxy XR’s Chrome: https://<Surface-LAN-IP>:8443 proxies straight to the Spark hub.

MCP tools

A typical voice-agent flow: upload_image (a frame from the XR camera), image_to_textured_mesh (queues a mesh job, for example TRELLIS.2), wait_for_job (polls until complete, minutes on a Spark), and optionally generate_rig with rig_mode=template for an avatar.

Worker config: xr-ai/agent-samples/3daigc-vlm-example/yaml/3daigc_vlm_example_worker.yaml, with daigc_mcp_url: http://localhost:8260.

VLM backend: model_backend: nim uses a hosted NVIDIA NIM (NGC_API_KEY); local runs an on-Spark vlm-server instead.

Relationship to this project’s own client

Layer	NVIDIA XR AI	This project
XR input	Voice plus passthrough camera, through the `xr-ai` hub	WebXR controllers, World Library, the VRM viewport
3D generation	MCP, into `3DAIGC-API`	Task Manager’s own REST calls, into the same API
Output	Job files under the DGX’s own `outputs/`	Download, viewport load, and an RP1 publish
Companion agent	An in-hub VLM agent	A separate companion-chat handoff (open work)

Troubleshooting

Symptom	Check
The stack exits on start	`curl -sf http://127.0.0.1:7842/api/v1/system/health`, then restart the API
The MCP probe fails	Port 8260 already in use; `bash …/mcp/scripts/run_http.sh`
The headset cannot reach `:8088`	Run `xr-spark-hub-proxy` on the Surface’s own `:8443`
A mesh job never finishes	`monitor_xr_ai_3daigc_stack.sh`; check the Redis queue and GPU logs under `3DAIGC-API/logs/`
NIM errors	Confirm `NGC_API_KEY` is set, or switch the worker to `model_backend: local` plus `HF_TOKEN`