Caveman launches a Mac app that runs three coding agents at once

A developer who pays for Claude, Codex and Cursor can now run the three in one window. Caveman launched as a Y Combinator Fall 2026 company on 13 November 2026. The founder, Julius Brussee, had caught attention with a markdown file in the last year. It told Claude to speak like a caveman to save tokens. The Hacker News post gave the file 120,000 GitHub stars. Caveman Code is the full product that came next.

The app runs on Mac and in the browser. It plugs into the subscriptions and API keys the developer already has. A top-level agent called "Chief" starts chats, answers their questions, and reports back. Pull-request review shows a one-line summary. It flags the one choice that needs a person. Each change has a before-after explanation next to the code. A cloud mode drives a desktop and a browser so the agents can check their own work. The app transfers running chats to the cloud when the lid closes, so a session finishes on its own.

The numbers Brussee publishes

Every figure here is Caveman's own, from Brussee's launch post on Y Combinator. On Terminal-Bench 4 with GPT-6.1 Sol, he reports 21 percent cheaper runs and 31 percent fewer tokens. Caveman and Codex both solved 90 of 192 runs on that pair. On Terminal-Bench 4 with Claude Opus 5.5, he reports 14 percent cheaper and 22 percent fewer tokens. Caveman solved 97 of 192 runs. Claude Code solved 83. On a day-long session across connected apps on Opus 5.5, he reports 76 percent cheaper and 83 percent fewer tokens. Caveman passed all 48 steps. Claude Code passed 46.

The runs were done at list price with no router, Brussee writes. He calls them his own runs. No third party has repeated them.

Why the harness matters more than the model

Brussee's own explanation is useful on its own. On Claude Opus 5.5, he writes, the model's output is less than a tenth of the cost. The prompt, the tools and the re-reads carry the rest. A harness decides how much the model has to re-read on every request. Caveman's claim is that a tighter harness cuts that work.

One afternoon to decide if it keeps a seat

A team that pays two or three AI subscriptions can test Caveman without a new vendor contract. The app is free with the team's own keys and logins. A paid tier covers extra features, Brussee writes. The test worth running is one day of work on the real code base. Put Caveman on one side. Put the current harness on the other. If the invoice for that day shrinks, keep it. If it does not, the cost was one afternoon.

Source

Caveman (F26), Y Combinator.