After spending a lot of time trying to hack around and customize Claude Code, I made the jump to pi. It took me a while to get comfortable with it. While I don’t agree with all of their design choices1, it is a pretty great piece of software and quite fun to play with. The built-in documentation also makes it very easy for most models to customize and add extensions.
Combined with some OpenRouter credits, it has let me quickly try different types of models, get more of a sense for how they behave2, and do all of that from an environment I already know.
By now, it has gotten me to a place where I can comfortably work on side projects that have been abandoned for years because I could not find the energy to deal with all of the minutiae. After a few years, just getting a project to deploy can take all the energy I had. In my testing, the models did well with this kind of work.
-
I disagree with their assessment around security. The whole “it can do whatever it wants to do, you just need to trust the repo” seems crazy to me. Probably, like me, most people spend their first days building some kind of sandboxing mechanism. This is pretty wasteful and hard to get right, meaning a lot of people will mess it up and end up exposing themselves accidentally. ↩
-
It seems to me that the GPT5.6 family is a little “lazy”, you basically have to force them to do something. The Claude 5 generation, on the other hand, is proactive and will go quite far (in the case of Fable, I’d argue often very much too far). ↩