How to create an on-demand RPM mirror

I would like to create an RPM repository for Fedora packages on my local network. Due to storage limitations, I want the repository to be empty initially and download packages once they are accessed. ### Background I work a lot with local VMs. Anytime I create a new VM and install Fedora, a lot of packages are downloaded from the internet, and most of the downloaded packages are the same. To speed up the process I would like the RPMs to be cached on a server located on the same network. Similar questions have been answered with a combination of createrepo & reposync. I do not like the reposync part, because I don't want to clone the whole repository up front when I need only some of the packages. ### Ideal Solution I would like the server on my local network to act as an RPM repository for my Fedora installations. It should pass-through the metadata from whatever is configured in /etc/yum.repo.d/*. The server should deliver the requested RPM if it is present in the local cache, or else download it and then deliver it. A less ambitious approach would be to configure a single RPM repository instead of https://mirrors.fedoraproject.org/ ... and just use an http proxy. ### Update: 02 Nov. 2015 I already have an nginx running on the network, so I played around with a combination of proxy_pass and proxy_cache. It kinda works, but IMHO it has more drawbacks than benefits: - a separate configuration for every repo configured in /etc/yum.repo.d/*. - can't use metadata from https://mirrors.fedoraproject.org/ because of alternate mirrors. I dropped the nginx thing and installed squid, as suggested in comments. squid works great for me. With the store_id_program configuration, I am even able to use the alternate mirrors and still hit the cache, no matter where the RPM came from originally.