At some point someone tried a toy direct integration, but the only actually supported way is via a Python library that wraps llama.cpp in an OpenAI compatible API endpoint.