LinkGrammar looks very interesting and useful. But I believe that its scope is somewhat smaller than NLTK. Aside from sentence parsing, NLTK also gives you a bunch of tools for document tokenization, statistical analysis, classification, and some easily accesible corpora to try it all out on. I'd almost call it a natural language prototyping environment; it's a huge time-saver to those dabbling in the field or who need to quickly experiment with different techniques.

LinkGrammar is not really a toolkit and it is not written in Python. It is written in C.

Yup. My bad. There are python bindings, though.