Interesting question. The issue here is that tensorflow models cannot be serialized through pickle as weight matrices are saved to external files. In theory you can build a wrapper around keras classifier (with tensorflow backend) to make it pickleable. It works by saving weight matrices to memory. You can have a look at https://pypi.python.org/pypi/keras-pickle-wrapper/1.0.3. This is untested, so let us know if it works.
Having said that, you can perfectly use tensorflow in a Python recipe or notebook, outside of the "Custom model" interface.