How Toknary uses tokens

You stay in the terminal. Toknary decides where the tokens go, so you are not choosing a model or counting each request. The rule is simple: the expensive attention is for the plan, the everyday model does the writing, and the noisy work goes somewhere cheaper.

Three kinds of work

Search, summaries, formatting, and boilerplate are noisy. They do not need the most careful model. Everyday coding, refactors, and tests are the middle of the job. A plan, and a last look when the middle has already failed, is the careful pass.

Why the split exists

A coding session rereads a lot. If every file list and every summary uses the careful model, most of the tokens buy work a smaller model can do. We keep the careful setting for the moment that decides the change. The rest of the session stays lighter.

The pieces are written up separately: noisy work on a smaller model, plan, then write, and reading less on purpose.