The Tools I Wasn’t Using Were Costing the Most
My self-hosted AI assistant was quietly burning money on every cold start, and the culprit was tools I never used.
Here’s how OpenClaw works under the hood: every tool the assistant can call ships its full JSON schema into the system prompt. Parameters, nested objects, enums, the works. On a fresh start, all of that gets written into the model’s prompt cache at the premium write rate.
I had eleven tools enabled that I never touch. A browser tool. Voice-call and text-to-speech. Video and music generation. A whole family of remote-file tools for a machine I don’t pair. Each one a fat schema, all of it billed on every cold start.
So I disabled them. One config line, one restart.
The “new” tokens on a cold start, the expensive ones, dropped from about 380,000 to under 1,000. The cheap cached portion ticked up by 22k, because the stable parts now sit on the cheap shelf and get read back at roughly a tenth of write price. I moved the load from the expensive column to the cheap one.
Then I did the same thing across two other servers I run for colleagues. Same eleven tools, same collapse. Nothing was lost; any tool flips back on in about five seconds if a real job ever needs it.
The lesson for anyone self-hosting an agent: audit your enabled tools. You’re paying rent on every one of them, whether you use it or not.
OpenClaw is a self-hosted AI assistant you run on your own server, with full control over its tools, memory, and cost.
I teach two classes on setting up and getting the most from OpenClaw on Udemy: Easy OpenClaw and Get Real Work Done With an AI Assistant.
