Skip to main content

​
How usage is measured

The agent reads each AI tool’s own local session history, read-only, and turns every model call (or, for tools that record only prompts, every prompt) into a usage event. Prompts and replies stay on the machine: only token counts, model, time and cost are kept. Nothing needs to change in the tool, and a call is never counted twice, even after the agent restarts. Cost is worked out one of two ways:
  • Tokens at catalog rates. Most tools record tokens per call. The agent prices them with the model’s rate from the pricing catalog. A model with no rate is counted as unpriced, and the total is then shown as a lower bound.
  • What the vendor billed. Some tools record what the vendor charged rather than tokens. The agent counts that amount as is.
Both kinds go into the same totals, so a machine’s and a person’s cost cover every tool.

​
Cursor

Cursor no longer records usage in its local history, so the agent reads it from Cursor’s own stop hook, which Cursor runs after every turn with that turn’s model and token counts. Wherever Cursor is installed, the agent adds one entry to Cursor’s hook file: the machine-wide file on managed installs, ~/.cursor/hooks.json otherwise. The entry only records: it never blocks or changes anything Cursor does, and the agent’s Live guard entry and your own hooks in the same file are left as they are. It is removed when the agent is uninstalled. Each turn is counted at the catalog rate for its model. Cursor does not include subagents’ tokens in the hook, and it bills some usage from plan allowances, so the cost shown is an estimate.

​
Tools that bill in credits

GitHub Copilot. Each prompt is counted at the GitHub AI credits it used, the same unit GitHub bills in. Copilot does not record output tokens per prompt, so its rows show cost without a token count. Code completions are not billed in credits and are not counted. Kiro. Each prompt is counted with the credits Kiro recorded for it. The billing period line (credits used of the plan’s limit, overage and reset date) is the account’s, as Kiro last fetched it: every machine signed in to the same account shows the same figures. In the Glass Bar, Billed in credits lists only the tools used in the selected window, and a row with no tokens and no cost is left out of the Providers, Models and Agents tables.

​
Model names

Tools name models in their own ways. Where a tool writes its own identifier, the agent shows the model’s plain name: A model the pricing catalog does not know is still counted: in tokens, and in cost when the vendor billed it.