Basically I want to input a PDB file of a protein which I suspect to have a binding site. I want the software either to generate a small peptide (or a number of peptide candidates) capable of binding this protein, or to find existing small peptides which are suspected to bind.
This kind of predictive work has not been done with much success, especially not in a way that you propose. If this were routinely possible, we'd be able to control bacterial toxins, rogue protein kinases, or any other protein for which the crystal structure is available. I'd venture to say that a person who solves this problem will be either very rich (if operating outside of academia) or very famous, very likely both.
There are docking programs that will do what you ask when both partners are known, and will allow a degree of molecular flexibility in all participants. Most modern docking programs can do this. There are programs (Rosetta and its derivatives) that can design a peptide to fit a binding site by sampling different backbones and side-chains, but this requires lots of unique domain expertise and is still not something that can be done well for any given protein. I suggest you look into the research done by David Baker and his students and postdocs.
I second Mensur. You may look for some denovo ligand design software which can build peptides for you.However I would first try tools from Baker lab as suggested.
If you're interested , one such old example is "Emplics", but may be of little or no value now.
