Hi Matt,
Yes this is possible, in general the approach is to expose a python API to your C++ library, and wrap it with theano so that PyMC3 model can understand it. The best place to start is: https://docs.pymc.io/notebooks/blackbox_external_likelihood.html