If you spend any real time in Claude Code, you’ve felt it: the context window fills up faster than you’d like. Long files, sprawling tool output, half-relevant docs — all of it competes for the same budget. Once that budget is gone, you’re either compacting, restarting, or watching the model lose the thread.
That’s where RTK comes in. It’s a small tool I’ve started leaning on to keep my Claude Code context lean, and it’s quietly become one of the highest-leverage things in my workflow.
Why bother saving tokens?
A few reasons it matters more than people think:
- Better answers. Smaller, more focused context means the model isn’t fishing through noise. The signal-to-noise ratio is what actually drives output quality.
- Longer sessions. Every token you save is a token you can spend later — on more code, more iteration, more back-and-forth before you hit a wall.
- Lower cost. Tokens aren’t free. On long agentic runs, the savings add up fast.
What RTK actually helps with
The top things I use it for:
- Trimming tool output before it hits context — big file reads and command output get summarized down to what matters, not the full dump.
- Reusing context across sessions — instead of re-feeding the same project background every time, RTK keeps a compact version on hand.
- Pruning stale turns — older parts of the conversation that no longer matter get cleaned out so the live working set stays sharp.
Installing it for Claude Code
Dead simple:
npm install -g rtk
rtk init
Then point Claude Code at it once and you’re done. From there it runs in the background.
What it looks like in practice
That’s it. Short post, but if you’re hitting context limits in Claude Code, RTK is worth ten minutes of your time.
