But that's moving the whole LLM agent into the cloud, which creates its own difficulties. Not really a solution to the local secrets problem.

Tailscale's new aperture also solves this elegantly: https://aperture.tailscale.com/