When you use Claude Code, Codex, or Cursor to write code every day, have you ever crunched the numbers?
Those seemingly "routine" commands—ls, grep, find, git status—consume LLM tokens every single time. The LLM doesn't know you just want a directory listing; it swallows the entire output as context, and the billing meter spins wildly.
What rtk does is simple and straightforward: Before a command reaches the LLM, it intercepts it, compresses it, and only sends back the parts the LLM actually needs.
What is rtk
rtk (rtk-ai/rtk) is a CLI proxy tool written in Rust. It boasts 51,300+ stars and 3,100+ forks.
It consists of just one file—a single binary compiled from Rust. Zero dependencies. Download and run.
How it works: rtk inserts itself as a proxy layer between your AI coding tools and the LLM API. When you run ls -la /some/deep/path, rtk intercepts the command's output, analyzes which information is truly valuable to the LLM, and then sends only a compressed summary.
How Much Can It Save
Official data: 60-90% token savings for common development commands.
This isn't magic. Its approach is essentially "information distillation":
- Dozens of lines of filenames from
ls→ compressed into a summary of key files and directory structure - Hundreds of lines of changes from
git diff→ extracts the core patterns of the changes - Massive match results from
grep→ deduplicated, grouped, and key context extracted
The LLM doesn't need to see a complete directory listing to understand the project structure. It only needs to know "this project has three main modules, each with its own test directory."
Technical Highlights
- Rust Implementation: Single binary, millisecond-level startup time, extremely low memory footprint
- Zero Dependencies: No need to install Python, Node.js, or any other runtime
- Transparent Proxy: Completely transparent to your AI coding tools; no configuration changes required
- OpenClaw Plugin: Already supports transparent execution rewriting for OpenClaw
- Hermes Integration: Recently added integration support for Hermes Agent
- 960 Commits: Highly active project iteration, with new PRs merged just yesterday
Real Value
The math is easy to calculate: if you spend 8 hours a day using AI coding tools and consume 500,000 tokens daily, rtk can help you cut that down to 50,000–200,000 tokens. Over a month, the money saved might be enough for a cup of coffee—or if your team is larger, it could cover a team meal.
But rtk's significance goes beyond saving money. It solves a deeper problem: the LLM's context window is wasted on a massive amount of useless information. When the token budget is filled up by the output of ls and grep, there's less context left for the truly important code.
By feeding the LLM more refined information, rtk, in a sense, also makes it "smarter."
Who Is It For
- Developers who heavily use AI coding tools
- Individuals and teams sensitive to API costs
- Rust enthusiasts who enjoy tinkering with toolchains
Not ideal for: Users who only occasionally use ChatGPT to write a few lines of code—the savings won't be noticeable.