RTK — CLI Proxy That Cuts LLM Token Usage by 60-90%

Introduction

RTK intercepts shell command output between your AI coding agent and the terminal, intelligently compressing verbose results so the LLM sees only what matters. Written in Rust as a single static binary with zero dependencies, it drops into any workflow without configuration changes.

What RTK Does

• Intercepts CLI output and compresses it before it reaches the LLM context window
• Reduces token consumption by 60-90% on common dev commands like git diff, find, and ls
• Ships as a single Rust binary with no runtime dependencies
• Works transparently with Claude Code, Cursor, Codex CLI, and other AI coding agents
• Preserves semantic meaning while stripping noise, whitespace, and repetitive patterns

Architecture Overview

RTK acts as a lightweight proxy layer that wraps shell commands. When the AI agent issues a command, RTK executes it, captures stdout/stderr, applies format-aware compression heuristics (different strategies for JSON, diffs, directory listings, logs), and returns the compressed output. The compression is deterministic and lossless in terms of actionable information.

Self-Hosting & Configuration

• Install with a one-line curl command or download the binary from GitHub releases
• No configuration file required; works out of the box with sensible defaults
• Set RTK_ENABLED=1 to enable proxy mode for all commands
• Customize compression levels via RTK_LEVEL environment variable (1-5)
• Supports allowlist and blocklist patterns for selective command interception

Key Features

• Zero-config single binary with sub-millisecond overhead per command
• Format-aware compression: understands git diffs, JSON, file trees, and log output
• Reduces AI coding costs proportionally to token savings
• Open source under MIT license with an active contributor community
• Compatible with any AI agent that shells out to the terminal

Comparison with Similar Tools

• Native context management — Built into some agents but lacks cross-tool portability
• Manual prompt trimming — Requires user intervention; RTK automates it
• Output piping with grep/sed — Loses structure; RTK preserves semantic content
• LLM context window expansion — Costs more tokens; RTK reduces the need

FAQ

Q: Does RTK change the meaning of command output? A: No. RTK strips formatting noise and repetitive patterns but preserves all actionable information the LLM needs to make decisions.

Q: Does it work with any AI coding agent? A: Yes, it works with any tool that executes shell commands, including Claude Code, Cursor, Codex CLI, and custom agent setups.

Q: What is the performance overhead? A: Sub-millisecond per command invocation. The Rust binary is compiled to native code with no garbage collection pauses.

Q: Is it open source? A: Yes, RTK is fully open source under the MIT license.

Sources

• https://github.com/rtk-ai/rtk
• https://rtk.sh