@theorytoe I used to run locally until torch dropped support for the cards I was using and I couldn't be fucked having a jet engine under my desk anymore. I'm on a pay-as-you-go plan for API access now but I still only use open weight models
These days I'm juggling 3 lowest tier subscriptions (z.ai for glm, gpt and claude) I was an idiot for a month paying per token through openrouter because I was paranoid about using claude code/codex harnesses (you can use pi.dev with z.ai at least)
I feel like I'm a noob still, totally missed the first Fable drop because I hadn't even figured out I should be using claude code. I was considering signing up for a 4th sub with MiniMax to check out M3. GLM-5.2 has been surprisingly good with what I've thrown at it so far, tempted to try it on more serious stuff, but it eats usage like crazy compared to the others
I am running local models for image gen, roleplay (with silly tavern) and even an AI companion with a LLM.
However, due to the limitations of the shit I can run locally (I'm limited to a 30B MoE model as a LLM with my current hardware, which is not very smart), I also pay for Claude to help me with more convoluted shit.
@harblinger ok so the tab thing is just a swap between a 'plan' and 'build' mode (i assume plan has no write access while build does) also seems to have free models or something (what ive been using, some deepseek type shit)
just from the quickest skim of opencode (havent tried it) I'd say the difference is pi.dev is barebones and you can add extensions as you please, I saw this in the opencode readme: >OpenCode includes two built-in agents you can switch between with the Tab key. so it sounds like it's just more opinionated on what works well, maybe I'll try it sometime since I sort of fell into that pattern in the past: discussion and writing a SPEC.md and then passing it off for implementation
@theorytoe@harblinger You can add custom agents and skill sets to opencode. It's sorta compatible with stuff that's made for claude code (which is the most popular), but it's not a drop-in