irvin.dev

OpenCode

I’ve been using Claude Code for over a year as my LLM harness of choice. Claude has been truly impressive and a real gateway into a new world of productivity for me at work and in my personal projects. But we are reaching a point at which the open weight models have caught up enough for my needs and I’d really like to move away from walled gardens for my LLM usage.

Plus, Claude has been incredibly slow lately. I had been on the Max x5 plan for quite a while, but for the last few months I’ve been back to Pro. And now, this is probably my last month. I’m considering a mix of OpenAI’s Plus plan and pay-as-you-go API usage through OpenCode Zen and OpenRouter. As for my reasoning for pivoting to OpenAI, they offer more generous usage and allow you to use the harness of your choosing. Ultimately, I’d like to be open weight API usage only. This is a step in that direction.

I’ve been playing with Skills in OpenCode, and I think I’m making my own app, Liminal Salt, redundant. I felt bad about it, about putting in all this work trying to create something that other people might use, only to discover tools like OpenCode can simply do it better. But I realized I’m just one person with a full-time job; I could never make something better than projects that have teams of developers focusing all their efforts.

OpenCode is great. Extensible. Open source. Private. And I can use it to accomplish everything I need in my LLM usage. Claude Code is a terrific first step, but it’s opinionated and locked down and becoming bloated and slow. I want to feel more in control.

The last year or so has been a whirlwind as these models get better and better. It’s an exciting time. While LLMs are still wrong a lot, proper attention and clear instructions and a good harness can really make them shine. Rarely do I code by hand anymore and in another couple years I imagine it will be even crazier for software development. No predictions, I’m just along for the ride at this point.

But I will say that I don’t see how companies like OpenAI and Anthropic can sustain the outsized investments they’re receiving. With open weight models like Kimi, GLM, Deepseek, and others, the competition is ever only six months behind. And for the general public, I just don’t see the utility of LLMs enough for these companies to earn back their spend. LLMs aren’t going away, they’re going to change a lot, but they’re setting a lot of money and resources up in flames.