kevincxy — notes on systems, infra, and AI

For the past year or so, my AI coding setup has been a bit of a split personality. At work, I live in Claude Code (Opus 4.8), and I'm a genuine power user. It runs in my terminal all day, and I rarely close the session. At home, though, I reach for something different: OpenCode Go, with minimax M3 as my default model.

That might sound surprising coming from someone who uses Claude all day, but after three months of running OpenCode as my personal driver, I genuinely don't miss Claude Code at home. Here's why.

The pricing math just works

Let's start with the obvious one. Claude Code is $20/month. OpenCode Go is $10/month, and your first month is just $5. For what I get out of it, that pricing is honestly hard to beat.

The generous usage on the Go tier is what sealed it for me. I run minimax M3 most of the time, which gets me roughly 3x the usage of the standard tier. In a typical month, I:

Build out small web app ideas
Revamp this very website using it
Manage parts of my homelab (more on that in a sec)
Tinker on side projects that may or may not ship

And even with all of that, my monthly usage rarely cracks 10%.

What I actually use it for

OpenCode isn't just a toy for me. It does real work.

Web app ideas to reality. I have a long list of app ideas that have been sitting in a notes app for years. With a $10 subscription that I don't have to think twice about, I can finally bring them to life at the rate they're coming to me. That's probably the biggest unlock for me personally.

This website. The redesign of my blog (yes, the one you're reading right now) was built with OpenCode. From layout tweaks to content edits, the agent handled the boring parts and let me focus on the writing.

Homelab management. I run openclaw and hermes, and OpenCode helps me manage both. Configuration changes, container debugging, new service deployments; it all goes through the terminal. Because OpenCode is just as comfortable in a TTY as it is in a fancy IDE, it slots right into my homelab workflow.

The Claude Code comparison

Here's the part I didn't expect to write: I don't really miss Claude Code at home.

As a daily power user of Opus 4.8 at work, I was worried OpenCode would feel like a downgrade. Slower, dumber, more frustrating. The opposite has been true. For the kind of work I do outside of work (smaller scope, fewer files in play, more iteration), the experience is genuinely comparable. OpenCode is a great competitor in this space, and the value proposition is impossible to ignore.

If you're already a Claude user, you're not going to be shocked by the interface or the workflows. It feels familiar.

Free vs paid models: a real difference

I also recently switched my openclaw and hermes agents over to use my OpenCode Go API key, with deepseek v4 flash as the underlying model. I had been running these on free-tier models before, and the difference is night and day.

The paid deepseek v4 flash is:

Noticeably faster, no more waiting around for tool calls to return
Less error-prone, fewer loops and fewer "I can't access that" moments
More confident, it commits to an approach and executes instead of asking me to confirm every step

If you have agents running in the background doing real work, the paid tier is well worth it. The free tier is great for learning, but the moment an agent becomes part of your actual workflow, you want the speed and reliability.

Final thoughts

OpenCode Go has become one of those subscriptions I forget I'm paying for, in the best way possible. It's just there when I need it, it gets out of my way, and it lets me ship ideas faster than I can come up with them.

If you've been curious about trying OpenCode, I'd love for you to give it a shot.

If you want to try it, here's my referral link. Clicking it gives you $5 in credits and gives me a small kickback too, so thanks for the support if you do. If not, no hard feelings at all; OpenCode is great either way.