Run the XR AI browser demo

Execute the quickstart document's Simple VLM example and return a working browser link, using setup when preparation is needed. Repository paths are relative to the selected checkout's root, not the installed skill directory.

Choose the environment

Use the checkout, repository ref, and model environment already selected by the user. If the ref is unknown, ask whether to use the latest release or main. Follow setup when the checkout, ref, or model environment is not selected, or prerequisites are unverified. Retrieve it from https://raw.githubusercontent.com/NVIDIA/xr-ai/<ref>/skills/setup/SKILL.md, using the selected ref; no separate installation is needed. Keep the skill, checkout, and docs on that ref. If this skill is absent at the selected release, ask before using main for all three. Read the selected checkout's AGENTS.md.

Default to the current host. If the user selects an existing remote host, including Brev, obtain any missing connection details and execute there. Provisioning a machine is a separate task. When needed, follow setup on that host to choose local or hosted model inference.

Prepare and launch

Before starting the sample, follow the ONNX Runtime telemetry guidance in docs/source/getting_started/networking.md. If outbound telemetry is unwanted, apply its environment opt-out before launching the process.

Before launch, follow docs/source/guides/customizing-model-servers.md for existing services, reuse requirements, replacement behavior, and shutdown scope. Reuse a compatible running demo; obtain permission before replacing existing services.

Execute the Simple VLM procedure in docs/source/getting_started/quickstart.md. For hosted inference, use the endpoint configuration validated during setup and skip local model startup.

Keep the sample alive in a managed terminal task. Record its directory, commands, GPU profile, stop handle, logs, and preexisting services, using reuse messages and startup logs as evidence. Wait for readiness per the quickstart. For failures, follow docs/source/guides/troubleshooting.md and the model-server guide, recording retained services. Report actionable blockers.

Connect and verify

Check the web-client URL from the browser's machine. For a remote host, follow docs/source/getting_started/networking.md for the browser address, TLS, and WebRTC media access. Obtain permission before changing firewall rules or broadening network exposure beyond the configured deployment.

Open the page when browser tooling is available. Follow the quickstart's connection and interaction steps, using a camera frame when media permissions allow it. Verify a connected client receives an agent reply. If browser interaction is unavailable, state that the conversation remains unverified.

Keep the handoff short: the browser link, remaining certificate and connection actions, one interaction from the quickstart, verification status, any configuration edits, and how to stop this run. Leave the demo running for the user. In the stop instructions, identify the recorded launcher task or PID and explain how to send it Ctrl+C (SIGINT) or SIGTERM so it cleans up its children when the user is finished.

Give model shutdown instructions from the model-server guide using the launch checkout and GPU profile, explaining their whole-stack scope. If services predate this run or ownership is uncertain, leave models running unless the user authorizes that shutdown.