Do you use coding agents within NixOS? If yes, what is your workflow? Which packages you use, services, etc?
I was hoping to prepare one using Ollama as backend. Do you have any suggestions?
2 Likes
I run the Pi coding agent in a arch distrobox, it seems to work well. There are also some third-party flakes (GitHub - lukasl-dev/pi-mono.nix: Nix flake for pi, a terminal coding agent. · GitHub) for it, but I prefer the distrobox as Pi moves much faster than I update my flake inputs, and I am also unsure how well they work with extensions. It should work with Ollama and home-manager services.ollama.
3 Likes
I’ve got my own research project and framework, and giving it full access to NixOS to rebuild its VM has been really funny to watch.
I’m currently starting to play around with litellm (very early, just started looking into it today), and for a backend I personally use lmstudio (but a friend swears by vllm).
I would really like to play around with local models a bit more, but I am currently using claude-code wrapped in bwrap. I hooked up a NixOS test which loads my NixOS configuration and the coding agent uses it interactively to run commands and take screenshots etc. This useful for validating runtime behaviour that cannot be caught at eval time. I am still learning and experimenting, but I think NixOS is uniquely positioned for this kind of workflow.
The following repository packages quite a few llm tools, which are updated more regularly than the ones in nixpkgs.
3 Likes