I just wanted to add one more clarification, because actually there is a bug in the code I posted.. For this late binding to work properly.. You really want to have something like this (sorry, ripping from non-test code): template<typename T> struct PAccessorConverter { static PyObject* convert(const T &x) { if (x==0) // i believe this is sending back None, not sure have to verify return boost::python::incref(boost::python::object().ptr()); return boost::python::incref(x->getPtr().ptr()); } }; and have that instantiated something like #define IMPLEMENT_PYTHON_CONVERTER(x) \ boost::python::to_python_converter<x *, PAccessorConverter<x *> >(); IMPLEMENT_PYTHON_CONVERTER(myclass) and then when you're doing a class.def, don't return the "existing_reference" because that will somehow cause boost python not to look up the converter.. instead do a return by value... something like: .def("getParentUser", &PCamera::getParentUser, return_value_policy<return_by_value>()) this return_value_policy<return_by_value> queues boost python to search the converters.. which will eventually hit the converter that uses the embedded python object.. (which allows you to have a 1-to-1 relationship between python and c++) -tim