Hi everyone,
I’m Nils, a solo founder from Belgium. I’m currently building my AI-agent setup from scratch, locally on a Fedora laptop with rootless Podman — no VPS yet, no big budget, no fancy infrastructure.
My goal is to build a practical AI stack for research, automation, and eventually cashflow, while keeping costs and complexity under control.
Small tip that may help others here: I recently added Caveman mode to my local OpenClaw/Hermes setup.
GitHub:
The project claims around ~75% reduction in output tokens by making the agent answer in a much shorter, direct style while keeping technical details intact. From my own early use, the biggest benefit is not just cost — it also reduces context mess and makes long sessions easier to manage.
Important nuance: as far as I understand it, this mostly reduces the agent’s output tokens. It does not magically reduce all input/context tokens.
There is also a more extreme option called Wenyan. It is a much more compressed language mode and can become classical-Chinese-like. I personally would not use it as default for human-facing replies, but it may be interesting for ultra-short summaries or later agent-to-agent compression. People should probably test and decide for themselves.
For my situation, Caveman is the practical default:
- fewer output tokens
- faster replies
- less context bloat
- better fit for a low-budget local setup
- less chance of drowning in long AI answers
I’m still learning and building, so I’m mostly here to observe, test, and share what actually works when starting from almost nothing.
Curious if anyone else here is experimenting with compression prompts, short agent languages, or token-saving workflows.