I built a tool www.lovespreadsheets.com earlier this year and a huge problem we kept running into is processing messy, real-world, complex spreadsheets for AI agents.
Even a "medium" sized spreadsheet converts to about 1m+ tokens if read fully, which immediately blows up the current context windows of most models. So I noticed most agents sample data from spreadsheets in chunks instead.
So I (& Claude) built this library Sheetgraph, https://github.com/asharma327/sheetgraph, to compress spreadsheets (~100x compression) while still maintaining their structure (like a dependency graph for formulas) & other relevant metadata such as macros, cell data types, formatting. There is a zipped full workbook of the spreadsheet as well for the agent to lookup data as needed.
Still rough around the edges so would love any feedback on both the idea and any details. This is used in our primary product + services platform Love Spreadsheets so we are actively maintaining and improving it.
Thanks!