Usage and Token Tracking
Jupiter keeps local token-usage history so you can answer a fairly basic question: which projects and models are actually consuming tokens?
When a model provider returns data, Jupiter stores input, output, total, cached-input, cache-write, and reasoning token counts.
It also keeps the operation type and model information. If an agent preference falls back before a request, usage is recorded against the model that actually handled the request.
Context-compaction usage is tracked separately so its overhead remains visible.
Local usage history is retained for 60 days.