It runs the real command, trims the repetitive parts, and leaves command behavior alone.
Two real receipts from actual work:
Java/Gradle task (Claude Code): 5,330,571 -> 90,127 estimated tokens across 88 commands (98.31% saved) Research task across 4 repositories (Claude Code): 944,007 -> 59,195 estimated tokens across 96 commands (93.73% saved) You can try it with the install script, run ccp init, then check the effect with ccp gain.
It is not a magic win on every command. In this repo's own usage with Codex, overall savings were lower when the mix was heavy on passthrough-heavy commands like sed and openspec.
It also works across a broad set of agents through ccp init, including Codex, Claude, Cline, Cursor, Gemini, GitHub Copilot, Roo Code, Aider, and others.
Repo: https://github.com/SuppieRK/ccp
If you use coding agents heavily, which commands waste the most context in your workflow?