AI spend you can see — checked before the turn, charged after it

AI features usually hide their cost until the invoice arrives, and they fail in the worst possible way when a budget runs out: mid-answer, with no explanation.

VibeControls wraps every assistant turn in two explicit steps. Before the message is dispatched, a pre-flight credit check runs; if the monthly budget is depleted the gateway rejects it and the panel blocks the send and shows you the upgrade path, rather than starting an answer it cannot finish. After the turn completes, the real token counts — prompt and completion — are reported and charged.

What is charged is computed on the server from the model's own token prices, never from a number the browser supplies, and the charge is keyed on the assistant message it belongs to. A retried or duplicated report therefore cannot double-bill you. Because the weighting is per model, a cheap model costs proportionally less than an expensive one instead of every turn costing the same flat amount.

What you see is a credit meter in the assistant panel that refreshes after each turn, and an explicit out-of-credits state with guidance when the budget is gone. Harness sessions have their own view: the per-agent AI stats tab reports session and usage counts straight from the agent, because those turns spend your own provider key and are not metered in credits at all.

Two honest notes. First, that split is the whole point — platform inference is metered, your own keys are not, and there is no path where VibeControls quietly bills you for tokens you already paid a provider for. Second, workspace-admin cost dashboards and anomalous-spend alerts are on the roadmap, not shipped: today the transparency is per user and per agent, not an org-wide spend view.