Show HN: Lowfat – pluggable CLI filter that saved 91.8% of my LLM tokens

A new Show HN project called Lowfat rolled in with a big promise: trim down the wall of text your command line spits out before it gets sent to an AI helper, and you could save a jaw-dropping 91.8% of your tokens. In plain English, it’s trying to stop people from paying to send mountains of boring computer output to chatbots. The creator pitches it as small, private, and easy to plug into your existing setup. Very tidy, very practical, very "look honey, I made the AI bill smaller."

But the comments? Instant trust issues. The biggest worry was brutally simple: what if the tool throws away the one line that actually mattered? One commenter basically summed up the whole drama as, “Congrats on saving money, but did you just delete the clue the AI needed to solve the problem?” Ouch. Another group wanted a real showdown with rival tools, especially since some of those already have a reputation for occasionally “helping” a little too hard and stripping out key details.

Then came the classic internet dogpile on documentation. People weren’t satisfied with being told how it works; they wanted receipts showing what it actually does in normal human terms. The funniest hot take was the tough-love version: forget fancy filtering, just block giant outputs entirely and force the AI to write its own filter like a student who forgot their homework. Even the naming debate got airtime, with one commenter wondering what on earth we’re supposed to call these tiny AI-behavior-tweaking gadgets. So yes, Lowfat arrived as a cost-cutting hero, but the crowd turned it into a referendum on trust, clarity, and whether “helpful” software is secretly a chaos gremlin.