// the find
rtk-ai/rtk
CLI proxy that reduces LLM token consumption by 60-90% on common dev commands. Single Rust binary, zero dependencies
RTK is a CLI proxy that sits between AI coding assistants and shell commands, filtering and compressing output before it hits the LLM context window. Single Rust binary, transparent hook integration, supports 100+ commands across 14 AI tools. Aimed at developers running agentic coding sessions who care about API costs.
The core architecture is right: most command output is noise (git verbosity, progress bars, ASCII art, redundant whitespace) and the shell layer is exactly where to cut it. The PreToolUse hook transparently rewrites commands before execution — the LLM doesn't need to know RTK exists, so you get 100% adoption without prompt engineering. The tee fallback is genuinely clever: when a command fails, it saves full unfiltered output to disk so the agent can read it without re-executing, preventing the filtering from swallowing debugging context at the worst moment. The custom TOML filter DSL means users can add filters for internal tools without waiting on upstream.
The savings numbers are self-reported estimates from unspecified 'medium-sized' projects with no reproducible methodology — the 80% figures in the table are plausible but unverifiable. Windows support is a real gap: the auto-rewrite hook doesn't work on native Windows, only falling back to CLAUDE.md injection, so the main selling point is absent unless you're in WSL. The filtering is inherently lossy — if a command exits 0 but the truncated output hid something important, the agent makes decisions on incomplete information with no warning. Built-in filters are compiled into the binary, so if a filter mishandles your version of a tool, you're waiting on a release rather than editing a config file.