Git knows about three things; blobs, trees, and commits. A blob is a collection of bytes. It is stored on disk as a file containing the bytes named after the sha1 hash of the bytes. A tree is a set of mappings from name to a blob or a tree, allowing you to represent collections of blobs like they exist on disk. Finally, the commit object is some metadata (author, committer, date, message, etc.) and a tree object. Add some shell or Perl scripts on top of this, and you have Git.

So if you can write a simple shell script, you can make git do whatever you want it to do.

(That's why git seems so "messy" from an implementation standpoint, it's a bunch of shell and perl scripts that do random things. Darcs and Mercurial have a slightly more complicated model, and hence everything has to happen "inside". This results in a cleaner looking interface, but a lot less hackability.)