If LLMs have models (tokens), there are definitely tokens inside said models, and a program can manipulate them.