Two things:

1. You may not want to expose bits and pieces of your data and metadata to an LLM, you dont want your data to be used for training. If you are using LLM running on your machine, as in this case, you are covered there.

2. Claude can do a lot of stuff, but doing multi step analysis consistently and reliably is not guaranteed due to the non-deterministic nature of LLMs. Every time it may take a different route. Nile local offers a bunch of data primitives like query, build-pipe, discover, etc. that reduces the non-determinism and bring reliability and transparency (how the answer was derrived) to the data analysis.