Aticaa.
rtk — an AI Token-Saving Tool, and My Setup Mistake

rtk — an AI Token-Saving Tool, and My Setup Mistake

· #dev · 한국어

rtk — Rust Token Killer

Rust Token Killer, or rtk, is a CLI proxy that puts command output on a diet for LLM dev sessions. Every time an agent runs git status, ls, or npm run build, the verbose output piles into the context window; routed through rtk, the same information arrives compressed — it claims 60–90% token savings on common dev commands. It’s a single Rust binary with zero dependencies, and it has passed 70k stars on GitHub.

The mechanism is elegant: hook it into Claude Code and the agent’s commands get transparently rewritten as rtk git status. The agent never knows its commands pass through a proxy. And rtk gain lets you measure what you actually saved — that measurement command is where today’s story begins.

Getting it running takes two commands:

brew install rtk   # install
rtk init -g        # wire the Claude Code hook — forget this and you get today's story

I installed it two weeks ago. Today I checked on it.

Total commands:    159
Tokens saved:      0 (0.0%)

[warn] No hook installed — run `rtk init -g`

Zero percent. It processed 159 commands without saving a single token. The reason is embarrassingly simple — I installed it but never ran rtk init -g. Without the hook, rtk intercepts nothing. The tool spent two weeks diligently doing nothing at all.

Installed is not working. A tool that claims an effect deserves one measurement. This case only came to light because rtk gain honestly reported zero.

The Innocent Suspect

The second story is the more embarrassing one.

That same day, Tia — our AI partner — kept hitting strange errors. ls threw an --icons option error; cat printed usage for some unfamiliar tool. Tia’s diagnosis: “rtk must be rewriting my commands.” From then on she tiptoed around it, routing around broken commands and noting “rtk ate it again” each time.

But as we just saw, rtk had no hook installed and was doing nothing. She spent the whole day interrogating a suspect with a perfect alibi.

The real culprit was ~/.zshrc:

alias ls="eza --icons"    # pretty icons
alias cat="bat"           # syntax highlighting
alias find="fd"           # intuitive search

Modern CLI aliases, added to make the terminal pretty for humans. Claude Code’s shell inherits the user profile, so when an agent speaks POSIX, eza answers in its own dialect. The -t in ls -lt (sort by time) got misread by eza as an --time option missing its value.

The Fix Is Two Lines

Just turn the aliases off for agent sessions. Claude Code sets a $CLAUDECODE environment variable:

if [ -z "$CLAUDECODE" ]; then
  alias ls="eza --icons"
  alias cat="bat"
  alias find="fd"
fi

Icons keep sparkling in the human terminal; the agent gets plain ls.

We also planted the rtk hook properly. A fun detail: in non-interactive mode, rtk init -g refused to patch our settings file on its own and printed manual instructions instead. A tool careful about touching other people’s config — that earns trust.

The post-repair checkup:

rtk checkup terminal — telemetry consent: no, savings starting at 0.1%, discover says “usage looks good”

With the hook biting, savings moved off zero to 0.1% (the first rtk grep saved 24.2%), and rtk discover returned “No missed savings found.” Telemetry remains declined — measurement stays local, the data stays ours.

What We Learned