AUDR by Chargebee

AUDR stands for Agent Usage Detail Record. It's a specification, not a dashboard: one JSON object per metered operation, with a schema that any agent harness, router, model provider or billing system can write and read. Chargebee drafted it and publishes the spec under Apache 2.0.

The gap it fills is attribution. Your app knows which customer triggered a run. The router knows the token counts. The tool layer knows the sandbox stayed open for four minutes. Each system is observable on its own, and none of them can tell you that Acme Corp's support agent cost 38 cents last Tuesday. Most teams feel this first as an agent cost tracking problem, because the monthly model invoice hides which feature burns the budget.

The design borrows from telecom. Carriers settled on the Call Detail Record so one phone call could be described in a common format no matter whose switch handled it. AUDR makes the same bet for agent runs: a shared run ID, one authoritative writer per field, and merge rules a sink can apply without guessing.

Scope limits matter here. AUDR describes usage, not the invoice. Rating, the step that turns tokens into a billable amount, stays outside the spec, and the sink only validates, deduplicates and stores records.

Getting Started

1. Install the core package for your stack, either pip install audr or npm install @openaudr/audr, then pick a sink: a local JSONL file for testing or Chargebee for billing.
2. Create a client with that sink and an emitter block that names your component, app and version.
3. Register an adapter for the runtime you already run, such as LiteLLM, the Vercel AI SDK, Mastra or NeMo Relay, or call client.record() yourself.
4. Pass attribution with each request, usually an account or subscription ID plus an environment, then read the JSONL output or the joined records in your billing system.