Undergryd

I've had Kimi and GLM bookmarked to test for the longest time Two open-weight models I kept meaning to actually run. Never did. Every time it meant standing up a server, pasting keys, and watching a bill I couldn't predict. It never made the top of the list. This week it did. I've been testing @Undergryd. It takes the infrastructure work out of running an agent. I deployed a Hermes agent. Switched the model from inside the app; no developer, Telegram works too. Now I've got Kimi and GLM running side by side against my current stack. No server. No Docker. No terminal. Credits load up front and carry a ceiling, so there's no surprise invoice at month end. The question I've sat on for a year: does an open-weight model hold up on the work I'm already paying frontier prices for? Now I can just watch it happen. If you run agents for clients, you know the tax. A box per client. Keys everywhere. A bill nobody saw coming. Undergryd is opening early access to a small group of founding operators. If you've felt that tax, this is the moment to get in while the group is still small. Want a look? Head to earlyaccess.undergryd.ai