Skip to main content

​
rtk (Rust Token Killer)

rtk is a CLI proxy that filters and compresses command output before it reaches an LLM’s context window. A coding agent runs git status, cat, cargo test, etc.; rtk intercepts the command, runs the real tool, and returns a compact version — cutting the tokens the agent must ingest. Single Rust binary, 100+ supported commands, self-reported <10ms overhead, Apache-2.0. Repo: rtk-ai/rtk; site: rtk-ai.app. Version at capture: 0.28.2 (README also references v0.37.2 hook changes — versioning in the source is internally inconsistent). Name collisions (disambiguate before installing/citing): another rtk = “Rust Type Kit” on crates.io; “RTK” also commonly means Redux Toolkit. This is the AI-token one — install via cargo install --git https://github.com/rtk-ai/rtk, not cargo install rtk.

​
How it works

Positioned between the agent and the shell: Claude → rtk → git instead of Claude → shell → git. Four compression strategies applied per command type:
  1. Smart filtering — strip comments, whitespace, boilerplate.
  2. Grouping — aggregate similar items (files by directory, errors by type).
  3. Truncation — keep relevant context, cut redundancy.
  4. Deduplication — collapse repeated log lines with counts.
Editing/reading verbosity is tunable: rtk read file.rs -l aggressive returns signatures only (strips bodies); rtk smart file.rs gives a 2-line heuristic summary.

​
Auto-rewrite hook (the intended mode)

rtk init -g installs a PreToolUse hook that transparently rewrites Bash tool calls (git status → rtk git status) before execution — the agent needs no awareness of rtk. Supports 15 agents via each one’s interception mechanism: Claude Code / Copilot (PreToolUse hook, native binary), Gemini CLI (BeforeTool), Codex (AGENTS.md instructions), Cursor/Windsurf/Cline/Kilo/Antigravity/Kimi (project rules files), OpenCode/OpenClaw/Pi (TS plugin), Hermes (Python plugin), Factory Droid (hook). Important scope limit: the hook only fires on Bash tool calls. Claude Code’s built-in Read/Grep/Glob tools bypass it, so they are not auto-compressed — to get rtk output there you must call rtk read/rtk grep/rtk find or shell equivalents explicitly. (This is a real gap for Claude Code, whose agents lean heavily on the native Read/Grep/Glob tools.)

​
tee fail-safe

On command failure, rtk saves the full unfiltered output to a tee log and points the agent at the path — so the LLM can read complete error detail without re-running the command. Configurable (failures / always / never).

​
Token-savings claims — verify

The README’s headline “60-90%” and its per-command table (e.g. git push −92%, cargo test −90%, session total ~118k → ~23.9k tokens, −80%) are self-reported estimates, explicitly qualified in-source as “based on medium-sized TypeScript/Rust projects; actual savings vary.” No independent benchmark. Treat as (claimed, unverified) per wiki citation rules — a vendor illustration of expected magnitude, not measured performance. rtk gain reports locally measured savings per install, which is stronger evidence than the README table but still self-instrumented. Telemetry is opt-in / disabled by default (GDPR-framed); collects aggregate counts and anonymized command names only, no source/args/paths.

​
Connections

  • Same “stateless CLI built for agents” niche as ketch — but rtk compresses existing tool output rather than adding a new research surface; complementary, not competing.
  • Attacks the tokens-per-tool-output cost, adjacent to the tokens-per-task framing in Context Compression Strategies. Distinct from conversation compaction: rtk shrinks each observation at the source, before it ever enters the context window.
  • Hook-based, cross-agent command interception is the same integration pattern surveyed for other tools in Agent Harness.