Token Saver: RTK + Caveman

Two agents set up to spend fewer tokens on the same work: compact command output going in, terse answers coming out, and a hard cap on the total.

What's inside

• Claude Code and Codex CLI agents, independent of each other.
• RTK (connected to Claude Code): its shell commands (git, test runners, ls, grep, builds) are rewritten through RTK, which returns a compact version of the output. Any failure falls back to the original command. RTK works live with Claude Code, Cursor, OpenCode and Kilo (the last three in dangerous mode only), not with Codex. Install RTK from its card with one click (downloaded from GitHub, checked by SHA-256).
• Caveman (connected to both): answer-style rules added to the agents' prompts, so replies drop filler and pleasantries. Level full; switch to lite (keeps grammar) or ultra on the card. Works with Claude Code, Codex and Qwen Code.
• Budget: a token cap for the whole workspace, 5,000,000 tokens to start. When it is crossed, agents are interrupted and automated prompts are blocked until you raise or clear the cap.

How to use

Import, then work in either agent as usual. The RTK and Caveman cards show what they did; remove an arrow to switch a saver off for one agent.