RTK: The Token Killer That Sits Between Your AI and the Terminal

Every time your AI agent runs a CLI command, it reads the full raw output — and that output is noisy. A simple git status returns 15 lines of boilerplate. A test failure from cargo test dumps 200+ lines. Over a 30-minute session, you're burning tens of thousands of tokens on noise the LLM doesn't need.

RTK (Rust Token Killer) is a single binary that intercepts that output, strips the noise, and hands back only what matters. Here's exactly how it works — and how to wire it up on Windows with Cursor.

The Problem: Raw CLI Output Is a Token Furnace

Consider what happens when an AI agent runs git status without RTK:

On branch main
Your branch is up to date with 'origin/main'.

Changes not staged for commit:
  (use "git add <file>..." to update staging area)
  (use "git restore <file>..." to discard changes in working directory)
        modified:   src/components/Header.tsx
        modified:   src/pages/index.tsx

Untracked files:
  (use "git add <file>..." to include in what will be committed)
        src/utils/helper.ts

no changes added to commit (use "git add" and/or "git commit -a")

~200 tokens. Half of it is git telling you how to use git. The AI doesn't need a tutorial — it needs the diff list.

With rtk git status:

M src/components/Header.tsx
M src/pages/index.tsx
? src/utils/helper.ts

~20 tokens. Same information, 90% smaller.

How RTK Works: The Proxy Pattern

RTK applies four strategies per command type:

• Smart Filtering — removes comments, help text, boilerplate
• Grouping — aggregates similar items (files by directory, errors by type)
• Truncation — keeps relevant context, cuts redundancy
• Deduplication — collapses repeated log lines with counts

The agent never sees the rewrite. It just receives a much shorter answer.

Real Savings Over a Session

Here's what a typical 30-minute Claude Code / Cursor session looks like with and without RTK:

That's not just cost savings — it's a context window that stays clean longer, meaning the AI retains more useful history and makes fewer mistakes from lost context.

Installation on Windows

RTK ships as a single .exe. No installer, no runtime, no dependencies.

Step 1: Download rtk-x86_64-pc-windows-msvc.zip from the releases page, extract the rtk.exe.

Step 2: Place it somewhere permanent, e.g. C:\tools\rtk\

Step 3: Add that folder to your system PATH (search "Environment Variables" → System Variables → Path → Edit → New).

Step 4: Open a new terminal and verify:

rtk --version
# rtk 0.31.0

rtk gain
# No tracking data yet. Run some rtk commands to start tracking savings.

Wiring It Up with Cursor on Windows

RTK's bash hook — the mechanism that automatically rewrites git status to rtk git status — requires Unix. On Windows, it silently falls back to --claude-md mode, which writes instructions to a markdown file but doesn't intercept anything automatically.

rtk init -g --agent cursor

# Output:
# [warn] Hook-based mode requires Unix (macOS/Linux).
#        Windows: use --claude-md mode for full injection.
#        Falling back to --claude-md mode.
# [ok] Created C:\Users\you\.claude\CLAUDE.md with rtk instructions

The hook file (rtk-rewrite.sh) gets created but won't execute on Windows. To make RTK actually work with Cursor, you need to tell the agent explicitly via Cursor Rules.

Go to Cursor Settings → Rules for AI and add:

When executing CLI commands, always use the rtk prefix:
- git status/diff/log/add/commit/push/pull → rtk git ...
- cat/head/tail → rtk read ...
- grep/rg → rtk grep ...
- ls → rtk ls ...
- cargo test/build/clippy → rtk cargo ...
- npm test / pytest / go test → rtk test ...
- When in doubt, prefix any CLI command with rtk.

This is a one-time setup. Cursor Agent will now use rtk across all sessions automatically — no bash hook needed.

Verify It's Working

Open a repo in Cursor, ask the agent to check git status. Watch the terminal panel — you should see rtk git status being called, not plain git status. Output will be the compressed 3-line version, not the 15-line boilerplate.

After a few commands, run:

rtk gain

You'll see a breakdown of tokens saved per command type, with running totals. Over a week of active development, the numbers get surprisingly large.

Worth It?

RTK is a zero-config, zero-dependency tool that does one thing well: it makes every CLI call cheaper for your AI agent. On Mac/Linux it's completely transparent via hooks. On Windows it needs a one-time Cursor Rule — a minor setup cost for a tool that then runs silently in the background indefinitely.

If you're doing serious AI-assisted development, it's one of the highest-ROI additions to your workflow.

→ github.com/rtk-ai/rtk