I canceled my Claude subscription. While it was cool to play around with—and super helpful at updating a number of my scripts—I couldn’t really justify the expense. It also, as many others have noted, had some interesting architectural decisions around my scripts, but I’m not sure I would have fared any better. Right now, the Python scripts I built to handle summarizing my Mastodon home timeline have a hodgepodge of configuration and command-line options. I’m putting together a framework to merge the command-line and configuration options, among some other odd architecture choices. That being said, it works—and it’s quite possibly a little more polished than anything I would have written by hand in the same amount of time.
I’ve also gotten qwen3.8 and devstral working with aider in my local environment. They were pretty effective at doing some of the low-hanging fruit tasks I’ve tasked them with. The main issue is how long it takes to run a local model, but I think it’s good experience to tinker with local models. For a lot of conceptual use-cases, I think experience with them is handy for understanding how these models can be used in real-world scenarios. I don’t really envision everything connecting to Claude, or Gemini, or ChatGPT—there too many potential failure states. What happens if there’s a network hiccup? What happens when they change the API? What happens when your credit card expires and you forgot to change it? And—what happens when they increase prices? Local models are pretty handy.
I tried several different local claud-code style solutions and really only got two to function remotely well: opencode and aider. Opencode runs *extremely* slow and every local model seems to get the tooling configuration weird. Aider is much slimmer and it’s proven to be quite effective. I’ve tasked it with updating my scripts—and by configuring larger context windows—it has done a pretty good job on the small tasks I’ve set it about doing. It’s still somewhat slow, but it’s also helpful to watch the thinking process. Watching an AI think through the prompt I gave, ask questions, then make odd choices has helped me better understand how to be more specific in what I ask, which also seems to make the final run go quicker.