Imagine three messages arriving at a customer support site. We want to classify each one so it can reach the right team.

One reports a login failure. One asks for a new export field. One says that they were billed twice. Are these a bug, a feature request, or a billing issue?

This is a focused task for a local AI model. It can put the GPU already in my laptop to useful work.

Miranda, my Grok Bot agent, runs on a remote virtual machine. Through Thirp, she can reach Ollama on my laptop without opening an inbound port.

Here are the commands Miranda ran. She lists the models on my laptop, then sends those same three notes to gemma3:4b. This is the call running on her virtual machine through the actual Thirp dial.

On my laptop, GPU utilization starts at zero. Then during the call, it reaches 52 percent.

And here are the labels the model returned: T31, bug; T32, feature request; T33, billing. Those labels could be used to route the messages to the appropriate teams.

The answer comes back to Miranda with the model and tool identified beside the results.

GPU Offload is a downloadable MCP pack for local agents and authorized cloud callers. It is the first app in the Thirp Marketplace. Request early access at thirp.net.