An MCP (Model Context Protocol) Server to provide LLM access to PUDL Data?

[zaneselvans]

A user recently asked:

Would it be possible for you guys to make or solicit an MCP (model context protocol) server that gives LLMs access to PUDL datasets? That way, if one points his or her LLM session to the PUDL MCP server, one could ask English language questions around power data and the model would be able to perform the appropriate queries on the PUDL dataset via the server. Maybe the cost of providing such a server is prohibitive though.

It seems like "Parquet files in an S3 bucket" has to be a pretty common data source, and I'd be surprised if someone else hadn't already created the necessary glue for pointing an LLM at that kind of data. A little searching turns up this MCP Parquet Driver from @CDataSoftware:

• https://cdn.cdata.com/help/HIK/mcp/pg_mcpinstall.htm
• https://github.com/CDataSoftware/parquet-mcp-server-by-cdata

Are there other open source options available? How hard would it be to set up? Are there server-side operational costs, or is mostly consuming user-side LLM tokens?

[zaneselvans]

Simon Willison suggests that text-to-sql doesn't really require MCP. Just giving the model the database schema, table & column descriptions as part of its prompt, and then asking natural language questions. He made a demo of a CLI version of this at PyCon 2025.

Since DuckDB is relatively new, current frontier models like Claude 4, ChatGPT-5, or Gemini 2.5 will probably do better than older models.

But I don't really know how people are using these tools. Are there a lot of folks for whom having something they can install in a desktop LLM application like Claude Code with a button that then allows them to "Ask PUDL a question" would be helpful? Is there a way for us to create a prefab prompt with all of the PUDL metadata that's hooked up to DuckDB pointed at s3://pudl.catalyst.coop/nightly/*.parquet?

The datapackage.json file that we produce for the Parquet dataset should include all of the table & column descriptions, filenames, and data types. Maybe just dropping that whole thing in the context would work? It's 1MB of JSON.

[zaneselvans]

And some words of caution on this whole strategy from the folks at @marimo-team

https://youtu.be/fDiylmtYlWM

[zaneselvans]

It looks like there's a recent DuckDB community extension that provides SQL-MCP integration too: https://github.com/teaguesterling/duckdb_mcp