shiploop compared
Each comparison quotes the other project from its own README, says what it measures, and says plainly where it is the better choice. Nobody is deciding between these tools to be entertained.
- Caveman Claude Code: What It Saves and What It Doesn't Caveman shortens what Claude Code writes. Here's what its own README says it saves, what it doesn't touch, and when shiploop is the better fit instead.
- Ralph Loop Claude Code The Ralph loop runs Claude Code in a while-true loop against a plan file. Here's why the core idea is right, and what an unmanaged loop costs you.
The three places a Claude Code bill actually gets smaller
Tools in this space are easy to mistake for competitors when they are working on different parts of the problem. There are three, and knowing which one a tool touches tells you more than any headline percentage.
What the model writes. Output tokens: the assistant's prose, not your code. A skill that tells the model to be terse works here. It is the smallest of the three on most real sessions, which is why the projects that do it honestly report a modest whole-session figure rather than their headline one.
What the model reads. Input tokens: tool results, file contents, logs, diffs, and the tool schema block that ships with every request. A compressing proxy works here, and so does trimming the tool list a headless worker is offered. This is usually the larger half.
How many turns carry it. Claude Code resends the whole conversation on every request, so a byte added on turn three is paid again on every turn after it. Clearing, delegating to a subagent, and running each unit of work in a disposable session all act here. This is the only lever that removes bytes from the accumulation instead of shrinking them, and it is the one most tools ignore. The token usage guide works through the mechanism.
The comparisons above say which of the three each project touches, quote it from its own README or post, and name the case where it is the better choice than shiploop. If you came here to find out whether you can run two of them together, the usual answer is yes, because they are rarely working on the same third.