Hi Product Hunt. I'm the person who built this. The short version of why: A 20-person firm came to us with documents that legally could not leave their network. Cloud AI was off the table. So they bought a big-memory Windows workstation, got llama.cpp running - and then discovered the hard part wasn't the model. It was that 19 other people needed to use it. So we built the layer that was missing: per-user keys, quotas, an audit trail, a concurrency cap set from a real load test, and enough self-healing that it survives a weekend unattended. Two things I'd especially like feedback on: 1. The concurrency number. We measured 8 as the throughput peak on our iGPU box. If you've load-tested a similar machine and got a different curve, I genuinely want to see it. 2. The honest-limitations section. We ship a guard script for slot deadlocks instead of fixing them properly. I'd rather know if that's a dealbreaker for you than hide it in the changelog. This is early and rough around the ops edges. Tell me where it breaks.

Desktop Model Gateway

Turn one Windows box into your team's private AI server

Turn one Windows box into your team's private AI server