Due to the claude -p billing changes (yes I know they rolled them back at the last minute but if you read the rest of this you'll realize why I'm not going back) I decided to move off of Anthropic for agentic homelab use and general AI usage.

You can see my first post about this just after I had set all this up. Clearly, I was excited - and I still have that same level of excitement - controlling my home server via remote agent is so much fun and so cool. I expanded to have agents run on schedules and send me custom email digests, code on a schedule for some of my projects, and more.

My homelab was all Claude native:

I first looked for a provider that had subscription based billing. I found the OpenCode Go subscription to fit my needs, with OpenCode Zen for PAYG spillover. What followed was clearly the model(s), offered by OpenCode Go. It was time for me to dive into the Chinese open weight models. OpenCode Go says all of their providers serving these models have a zero retention policy, but I'm not sure I believe that. The price is great though. After a couple of weeks I've gotten a good feel for the Chinese models.

Then came the agent. I first started with OpenCode, which I enjoyed. They have a native opencode web command that launches a web UI - great for remote control. Connected it to a gated access cloudflare tunnel and my remote control was solved.

Last came the chat. I found Open WebUI. OpenCode Go's API key is OpenAI-Compatible which it supports. I signed up for a Brave Search API key and voilà, fully featured AI chat powered by open weight Chinese models. This also went behind a gated access CF tunnel.

Once my algorithms shifted because of my net activity around this migration, I eventually found the Pi agent. Pi is fantastic. I keep calling it the vim/neovim of agents.

However, this meant I had to move off of opencode web too. Well, good news, Pi has a fantastic developer community and pi-web exists.

What I learned

Every piece of the stack matters. Agent, Model, Provider each have their own downfalls, and "only being as strong as your weakest link" truly applies here. Any one of: agent polluting context, model just being bad, and provider being slow or over-quantizing (or anything else) can make your setup suck. Using Anthropic monolithically, I didn't think this way.

I love Pi. I highly recommend everyone try it and just see what you feel. If you love the Arch/neovim philosophy of give you just what you minimally need to run - then you as a user decide what features get added - this is the agent version of that.

China is doing a good job, even if it's built off of a backbone of distillation. You can't just distill to get where they're at. Deepseek has surprised me the most - I thought the "Deepseek Moment" was just an "ok China is participating now". Deepseek v4 Flash is genuinely a Google search replacement. Give it a Brave API key and it gives you better results, faster, for absolute dirt cheap. Deepseek v4 Pro is a good daily driver for agentic homelab usage as well, and for its output level it's also dirt cheap. If you want to code confidently, Qwen 3.7 Max and GLM 5.2 (!!!) are fantastic. Kimi, Minimax, and MiMo also have great models that deserve a try.