Most AI cost talk still focuses on the prompt. That is only part of the bill. What kept burning tokens for me was everything around the prompt: old context I should have trimmed tool output I kept dragging forward retrieval chunks that stopped being useful 20 minutes ago switching to a bigger model before the task actually needed it The annoying part is that none of this feels expensive in the moment. A session just gets a little messier. A little slower. A little harder to reason about. Then the token count quietly runs up. That is why I built TokenBar. It sits in the macOS menu bar and shows live token usage while I work with LLMs. Not after the session. During it. That changes behavior faster than a dashboard ever did for me. When I can see token usage climbing in real time, I am more likely to: cut dead context restart a bloated thread stay on a smaller model longer stop carrying tool traces that are no longer helping For me, the first win was not even cost. It was cleaner AI workflows.…