Half the tokens. Nearly the same work.
A terminal coding agent. On 25 matched gpt-5.5 tasks it spent 524k fresh tokens where Codex CLI spent 1,010k — 1.93× fewer — and passed 14 of the 25 to Codex CLI's 15.
Frozen since August 2026 · still installable and working
MIT · TypeScript · JuliusBrussee/caveman-code
A coding agent keeps paying for what it already read. So four layers sat under the loop.
Every turn, the same files go back up the wire. The same tool schemas, the same diffs, the same directory listing. The model is not doing more work — the transcript is just getting longer, and you pay for the length.
Caveman Code cut the resend in four places, then measured what was left.
- 01Caveman Modeoutput written compressed
- 02Tool budgetsa cap per tool, per turn
- 03Read-dedupa file is read once
- 04RTKoptional, off by default
Twenty-five matched tasks, run both ways. Including the one we lost.
Fresh tokens across 25 matched tasks on gpt-5.5. Caveman Code passed one task fewer than Codex CLI; both figures are printed above.
What shipped in the box. All of it still runs.
- 01
Autonomous goal loop with autopilot, and a read-only plan mode when you want to look first.
- 02
20+ providers: Claude, ChatGPT, Copilot, Gemini, Vertex, OpenAI, Azure, Groq, DeepSeek.
- 03
Architect and editor run as separate models, so the thinking and the typing are priced apart.
- 04
Session branching and shadow-git checkpoints: take a wrong turn, walk back to a known state.
- 05
Persistent memory through cavemem, plus MCP servers for everything else.
- 06
Up to seven subagents, each isolated in its own worktree.
Frozen, not deleted.
npm still has it.
Frozen since August 2026 · still installable and working
No longer in active development. Its compression ideas continue in Pebble.