Checking your usage
Where to see what you have used and what a run cost: in the app, on the web, and from the terminal.
WARNING
Harmony-routed model credits and a weekly inference allowance are not for sale, so there is nothing to meter yet. This describes a planned gateway product. It is not Harmony Cloud.
Three places, all showing the same numbers from the same source.
In the app
Settings shows the weekly allowance, how much is used, when it resets, and your purchased credit balance. The percentage is text as well as a bar, so it does not depend on distinguishing a colour.
On the web
Your account page shows the same figures plus a per-run breakdown: which model, how many tokens, and what it cost.
The breakdown is grouped by model, most expensive first, because "which model is costing me money" is usually the actual question.
From the terminal
harmony usage
For scripts, use JSON rather than parsing the human output:
harmony usage --json
The JSON shape is a compatibility promise. Fields may be added; the existing ones keep their meaning, because deploy scripts depend on them.
Note that an exhausted allowance exits 0. It is a normal state you chose,
and a non-zero exit would fail CI pipelines over it. Non-zero means we could
not answer the question: 2 if the device is not linked, 3 if the gateway
is unreachable.
Reading the numbers
A run showing "pending" is not free. It means the provider has not reported its usage yet. It will be priced within a day, and the total will rise. We show "pending" rather than zero specifically so the later correction does not look like a surprise charge.
If the itemised rows do not add up to what you were charged, the difference appears as its own line rather than being folded into a figure. That means we know the breakdown is incomplete and are investigating; the charge is the ledger figure. A breakdown that quietly rounded to match would be worse than one that admits a gap.
What is recorded
Model, token counts, credits, and time. Not your prompts, the model's output, or your file contents — those are not stored, which is also why support cannot show them to you if you ask.