19, Netherlands. I build Caveman: one compression engine, dropped in at every depth of your stack, each one cutting what your AI traffic costs. The top ones are free and open. The bottom ones report the savings in dollars you can audit.
| Caveman Skill | 65% fewer output tokens out of your agent. Code and errors stay byte-for-byte exact. | live, MIT |
| Caveman Engine | One command wraps the whole machine. Local compression, every original recoverable. | live, BSL-1.1 |
| Caveman Browse | Browser automation your agent reads in tokens, not pages. Compressed a11y snapshots, byte-exact recovery. | live, BSL-1.1 |
| Caveman Agent SDK | Agents that compress by default. Drops into Vercel AI SDK, Eve, Mastra. | v1 live, MIT · v2 in dev |
| Caveman Gateway | Point traffic at one URL. Caching, compression and routing, all eval-gated. | in dev |
| Caveman Enterprise | Same stack inside your own cloud. Verified savings, signed receipts, OEM embed. | in dev |
Labs: cavegemma — the compression baked into Gemma's weights. No system prompt needed.
Gateway, Enterprise, and the Agent SDK v2 are still being built. That's where most of my time goes.
Earlier experiments, frozen in August 2026. Still installable, still work — no active development. Their best ideas moved into caveman.
- caveman-code — terminal coding agent. 1.93× fewer tokens than Codex on the same tasks.
- cavemem — one memory shared across Claude Code, Cursor and Codex.
- cavekit — spec to plan to built software, with adversarial review.
Revu — spaced repetition for macOS. I built it because I wanted it for my own studying. Open-source core, product at revu.cards.
Pitchr — AI pitch coach. Built at HackEurope 2026 with @LucasDuys, @aravmdn and @Movenanter.
Stacklink — RAG over Google Drive that answers with citations. Closed.
Swift, Go, TypeScript, Python.
caveman.so · X · LinkedIn · Substack