Generating a good cli isn't all that hard for agentic coding tools. When you do it manually it's highly repetitive work. But all you are doing is low level plumbing. Given some parsed arguments, call a function, return the result (with some formatting, prettying, etc.). In the end it's just a facade for an API, library, or whatever else you want to have a cli for. Easy to write. Easy to test. But manually going through your API resource by resource, parameter by parameter, etc. takes a long time. An LLM just blazes through that in a few minutes. Generate some tests, tweak as needed, and you are good to go.

I did a few CLIs with codex in the last few weeks. I do simple ops with this stuff. I've had a few use cases for new features where previously I would have had to build some kind of quick and dirty admin UI just to use and test a new API feature before being able to integrate it into our product. With a generated cli, I can just play with it from the command line. Or make codex do that for me.

A good cli with a modern command line argument parser, well documented options, bash/zsh auto complete, pretty colors, etc. is generally nice to have. I mapped resources to commands and sub commands, made it add parameters with sensible defaults or optional ones. Then I got lazy and just asked it what else it thought it was missing, it made some suggestions and I gave it the thumbs up and it all got added. I even generated a simple interactive TUI at some point. Because why not? I also made it generate a md skill file explaining how to use the cli that you can just drop in your skills directory.