Did you train a self-hosted/open source LLM on your existing software and documentation? That should make it far more useful. It's not claude code, but some of those models are 80% there. In 6 months they'll be today's claude code.