Feature idea: local inference sensors

[jordimas]

Has this issue been covered in the Wiki?

• I have read the Wiki and my issue is not covered.

Is there an existing issue reported already?

• I have searched the existing issues, it does not exist.

Describe the new feature you would like

Would you consider optional sensors for local LLM servers?

Many GNOME users run local AI through Ollama, llama.cpp, LM Studio, or vLLM.
These expose localhost APIs or /metrics, but there is no simple top-bar view for idle / generating / slow / offline state.

Panel examples: AI 42 tok/s, Ctx 6.2k/32k, KV 72%

Initial localhost-only backends:

• llama.cpp: :8080/metrics
• Ollama: :11434/api/ps
• LM Studio: :1234/api/v0/models
  vLLM: :8000/metrics

This would be opt-in, read-only, localhost-only, and would not inspect prompts or completions.

I would be happy to prototype it if this fits Vitals’ scope.

[corecoding]

One of the "selling" points of Vitals is that it shouldn't cause stutter. One of the ways to prevent stutter is to prevent command line spawning. Fast forward, I had a ton of interest in GPU metrics which requires launching of nvidia-smi. While this works okay, it still pains me.

I've been slowly working on an idea where users can have Vitals run shell commands of their choice. This makes it so I don't get complaints about stutter, or if I do, I can simply say, don't do that. lol. By allowing custom shell commands, you could in theory spawn curl and have it return JSON. Or, if demand warranted, I could add an option to call custom endpoints and parse JSON.

Would this idea enable you to do what you want?