A CLI filter claims 91.8% token savings — the thread does the math

A demo GIF, a Rust binary, and one very round number

Lowfat is a small, pluggable CLI filter — about 94% Rust by the repo's own breakdown — that sits between your shell and your AI coding agent and trims unnecessary output before it burns context. It hooks into Claude Code via .claude/settings.json, ships an OpenCode plugin, or runs by simply prefixing a command. The selling points are modest and UNIX-flavored: "Small single binary," "Local-first — No telemetry," and "Composable — UNIX-style pipes… not magic."

The idea is real and uncontroversial: agents default to the broadest possible command, then pay for every byte of output. A git diff or a kubectl get -o yaml floods the context window with formatting the model doesn't need. Strip it first and you pay for signal, not boilerplate.

The 91.8% is doing a lot of load-bearing work

The number in the title is what carried the post — and what the thread refused to take at face value. The sharpest correction came from a developer who reverse-engineered exactly what was measured:

"It doesn't save 91% of tokens: it reduced in one user case 91% of output tokens on the raw CLI output… Once you apply Amdahl's law, there is no way this saves 91% of tokens holistically, which the title implies." — an engineer dissecting the benchmark claim

That's the whole gap: 91.8% of the output of one command in one case is not 91.8% of a session's tokens. The system prompt, the conversation history, the model's own reasoning tokens — none of that is touched.

A deeper worry was about quality, not accounting. Stripped output may be out-of-distribution for the model:

"LLMs were trained on the typical full-fat output found everywhere on the internet, and all of a sudden they get a slightly different response that may look like nothing they have seen before. Does that really save tokens in the long run?" — a commenter skeptical of filtering's effect on reasoning

And people who've shipped similar tools warned the failure mode is the agent fighting the filter:

"I've tried rtx and lean-ctx and these tools seem to end up confusing the agent more than helping. Any saving is irrelevant if the agent decides to work around the tool and makes even more calls than it would otherwise." — a developer who tried competing context-trimmers

The cleaner fix might be upstream, not downstream

The most pointed critique wasn't of lowfat but of the whole post-hoc category. Filtering after the fact still pays for the round trip:

"The bigger problem is agents defaulting to the broadest command possible. kubectl get -o yaml when a jsonpath query would give 1/50th the tokens… better to teach the agent to ask narrow questions in the first place." — a commenter who builds agent harnesses

Lowfat is a genuinely useful little Unix tool with a viral headline bolted on. The tool deserves the front page; the 91.8% deserves the asterisk the thread spent its energy attaching. The real lesson is the one the skeptics keep repeating: a percentage with no harness, no sample size, and no holistic baseline is marketing, not a benchmark — and "saved 91.8% of my LLM tokens" measured one command's output, not anyone's bill.