I settled on gptel, which is an LLM package for Emacs. It's kind of a mix of both approaches:

1. You have chats right there in the editor, easy to copy/paste and manage context/history without switching to a browser. You can also quickly add files or portions of files to the context or remove them again.

2. You can choose which model you want to use for what, granted you have an API key.

3. You can quickly highlight some code and ask for a change to it, which along with managed context is powerful.

I tried auto complete again and again but it doesn't work for me. At first I think "yeah, that's what I wanted to write", and then I have to look closer to realise it's not, and that completely breaks my flow and concentration. I can always write some pseudo code and proactively convert it to real code, I like to be in the driver seat.

Context management is really central to my workflow, I manage it like a hawk. Models tend to deteriorate as context content increases, in my experience, so I really try to keep it narrow.

For that reason, and because our clients didn't sign up for their code to be sent to Anthropic et al, I _mostly_ use models like I would use StackOverflow, not to generate non-trivial code I'd actually use.

But having the chats in my editor is really invaluable for me. Powerful text wrangling features make a difference in both speed and motivation.

I use it pretty heavily with pretty much only the high-end models and pay about $15 per month.