DGX Spark + Hermes CLI + Live Audio + Tailscale: My Local Alternative to Using Telegram

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Hi everyone,

I built Hermes Live Chat, a lightweight voice and text interface for accessing Hermes CLI on my DGX Spark from my phone. I really wanted an easy way to talk to my local agent while driving and I made some car safe features.

Connection:

Phone → Tailscale → DGX Spark → Hermes CLI → local model

It supports push-to-talk, live calls, wake-word mode, local speech recognition and TTS, attachments, and persistent conversations. Tailscale provides private HTTPS access without exposing the app publicly or using Telegram as the middleman.

Hermes Live Chat responding to a request

image
image1265×712 42.3 KB

Current resource usage:

  • Hermes Live Chat service: approximately 54 MiB RAM, with a measured peak around 277 MiB
  • Kokoro TTS: approximately 2.4 GiB RAM and 1.74 GiB GPU memory
  • Current DeepSeek backend: approximately190.5 GiB/unified memory

I’m currently using DeepSeek as the backend, and responses have been pleasantly quick. The interface is model-agnostic, so Hermes can be switched to another model later.

Screenshots, a mobile demo, audio sample, and installation instructions:

I’d curious to hear how other DGX Spark owners are remotely accessing their local agents.

2

N1ce, great idea :)

3

Recently integrated the same for my Opencode-based tooling. Frontend accessed over Wireguard to keep Telegram/Discord/whatever out. Now able to walk away from the desk while swapping between sessions. Can also use voice input for the TUI which Opencode doesn’t support.