RTK — CLI Proxy That Cuts LLM Token Usage

Introduction

RTK sits between your AI coding agent and the shell, intercepting command output and applying format-aware compression so the LLM receives fewer tokens. It ships as a single Rust binary with zero runtime dependencies, making it trivial to drop into any agentic workflow.

What RTK Does

• Intercepts stdout/stderr of arbitrary shell commands and rewrites them into a compact representation
• Applies format-specific reducers for JSON, diffs, directory listings, logs, and build output
• Preserves semantic content while stripping redundant whitespace, ANSI codes, and boilerplate
• Works transparently with Claude Code, Codex CLI, and other agentic tools
• Outputs original content when piped to a non-LLM consumer via TTY detection

Architecture Overview

RTK spawns the target command as a child process, captures its output streams, and passes them through a pipeline of format detectors and compressors. Each compressor is a pure function that pattern-matches on the input structure. The final output is emitted to stdout so the calling agent reads the reduced version. Because it is a thin wrapper with no daemon, startup overhead is sub-millisecond.

Self-Hosting & Configuration

• Install from source with cargo install rtk or grab a static binary from the GitHub releases page
• No configuration file required — sensible defaults cover most use cases
• Set RTK_PASSTHROUGH=1 to disable compression for debugging
• Custom reducers can be added via a TOML config at ~/.config/rtk/config.toml
• Integrates with Claude Code by prefixing shell commands or aliasing common tools

Key Features

• Single static binary, no Python or Node dependency
• Sub-millisecond startup; negligible overhead on command execution
• Format-aware compression preserves meaning while dramatically shrinking output
• TTY detection prevents accidental compression of human-facing output
• Open source under MIT license with active community contributions

Comparison with Similar Tools

• LLMLingua — prompt compression at the model level; RTK compresses at the shell output level before the prompt is built
• Repomix — packs entire repos into a single file for context; RTK focuses on individual command output
• Context window managers — most operate on chat history; RTK targets raw tool output

FAQ

Q: Does RTK change the command's behavior? A: No. RTK only post-processes stdout/stderr. The underlying command runs exactly as it would without RTK.

Q: Which AI agents does it work with? A: Any agent that shells out to run commands — Claude Code, Codex CLI, OpenCode, and custom setups.

Q: Can I disable it for specific commands? A: Yes, set RTK_PASSTHROUGH=1 or use the --raw flag.

Q: Is it safe for production CI pipelines? A: RTK is designed for interactive agentic use. In CI, run commands directly without the wrapper.

Sources

• https://github.com/rtk-ai/rtk
• https://rtk.dev/docs