I noticed that the startup time even when just running markitdown --help can be substantial (over 1s on my system).

This can be improved by avoiding top-level imports. It's quite simple, but requires to split the package, so I haven't submited a PR, as I realize it's a major decision for you to make. You can see my proof-of-concept here: AdrianVollmer@d57b959

I'd also suggest to add argcomplete for shell completion. Example in the fish shell:

What do you think?