Anthropic launched Claude Tag, and no matter how cliche it sounds, it actually seems like a shift in how we interact with AI. I also think that an agent acting across a team's data is only as good as the data it can reach.

Smoothie's idea is to treat data like Java treats source. javac compiles code once into bytecode that the JVM runs; Smoothie compiles a folder of mixed data (PDFs, spreadsheets, decks, recordings) once into a single artifact, and a small Rust runtime with no model in it serves it deterministically as a CLI tool. Every fact keeps a reference back to its exact source span, so when an agent answers it can point at the source line instead of guessing. Next step is exposing it over MCP so something like Claude Tag can query it directly.

The input processors are extendable: each is a small script, so you can write your own in any language, and a source can be a local file or a hosted tool.