rtk: Rust CLI Proxy Cuts LLM Token Consumption 60–90% on Common Dev Commands

The most under-hyped tool of the week is a single Rust binary with zero dependencies that sits between your coding agent and your terminal, compressing command outputs before they hit the context window.

rtk (Rust Token Killer) intercepts common dev commands — cargo test, git status, npm run build — and compresses their output by 60–90% before passing results to Claude Code, Cursor, Codex, or any other coding agent. A cargo test that normally dumps 155 lines gets compressed to 3. A git status shrinks from 119 characters to 28. A viral post circulating across KiloCode and CLIProxyAPI communities reports saving 10 million tokens over two weeks by adding a Claude Code hook that transparently rewrites commands to rtk-prefixed equivalents. GitHub

At 8,624 stars and climbing, rtk represents the highest-leverage single install for anyone running a coding agent. The tradeoff is real — compressed output can increase downstream generation tokens by ~18% in some workflows because the model has less context to work with. But the net savings are overwhelming for the typical agent session where terminal output is the primary context consumer.

The broader pattern matters more than the tool: the bottleneck in coding agent performance is increasingly context management, not model capability. rtk, the PostCompact hook in Claude Code v2.1.76, worktree.sparsePaths, on-demand MCP tool loading — they're all solving the same problem from different angles. Your agent's context window is the most expensive resource in your workflow, and most of it is being wasted on build output nobody reads.