Cost attribution
What AI agents cost, per task and per developer, with budgets and FOCUS export.
Every call has an owner
The loopback proxy counts each model call's tokens and ties the call to its task. The local dashboard shows spend by task and model, priced from a versioned catalog of models and prices rather than numbers in the code.
Accounting is as strong as the path it sees: where the sandbox makes the proxy unavoidable it is controlled (B), elsewhere an agent can route around it, and the figures are observed (C).
Budgets
A budget warns at a threshold, and a task that exceeds it can be suspended: every further decision is denied. By default a task and the subagents it starts share one budget, so a swarm cannot race past it: at most one last call can go over.
Loop detection notices an agent repeating the same action and stops it, and swarm limits cap how deep and how many subagents a task may start.
Across a fleet
Reconciling against the providers' own usage reports preview: Anthropic, OpenAI, Cursor, GitHub Copilot and Devin. Cost by developer and credential in Central Lite preview, and export in FOCUS, the FinOps Foundation's open billing format preview.
