AUDR by Chargebee is an open specification and instrumentation toolkit that records agent usage, raw cost drivers, and business attribution across distributed runtimes.
What the product does and how it is positioned
AUDR, short for Agent Usage Detail Record, is an open standard designed to capture end-to-end telemetry and cost attribution for AI agent workflows across distributed execution layers. When an agent touches application harnesses, model routers, and tool execution environments, AUDR links raw computational consumption—such as token counts, tool calls, and compute duration—directly to business metadata including customer accounts, features, and operational environments.
The standard operates on three core rules: minting a shared run identifier across every participating layer, maintaining clear authority per field so each fact originates from one authoritative source, and enforcing strict merge rules at the destination sink. Supported by core libraries in Python and TypeScript, AUDR emits records asynchronously and out of band, without inspecting prompt text or generated outputs.
Source-supported ways to use the product
Correlates resource consumption across routers and external tool executions with customer and feature identifiers supplied by the application harness.
Delivers structured usage records directly into local storage, data warehouses, or billing endpoints without imposing predefined pricing logic.
The documented workflow, where available
Install an ecosystem adapter or initialize the core SDK to monitor model completions, embeddings, tool scopes, and rerank calls.
Mint a shared run ID in the harness and attach runtime metadata such as customer identifiers and environment tags to outbound requests.
Allow routers and intermediate execution layers to measure token counts and compute duration while preserving the shared run ID.
Merge matching run and span data at the designated sink destination, enforcing field authority rules and delivering output to files or ingestion endpoints.
AUDR adapts the concept of telecommunication Call Detail Records to address data fragmentation in multi-tier AI agent architectures. An application harness creates a shared run identifier and passes it downstream in request metadata. Because intermediate routers and tools echo the identifier back, independent components can report operational metrics without losing the root context of the execution.
Data integrity is maintained through unambiguous field ownership and strict merge behavior. The application harness holds exclusive authority over attribution fields such as customer IDs and environments, while routing layers hold exclusive authority over token counts and provider metrics. Sinks assemble these records based on run and span IDs, rejecting conflicting submissions and treating any updates as separate correction records rather than in-place mutations.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
AUDR standardizes how agent execution metrics and attribution metadata are recorded across multiple application layers, routers, and tools.
AUDR builds on OpenTelemetry GenAI semantic conventions and can emit spans directly to OpenTelemetry collectors while adding durable record delivery semantics.
No external billing platform or hosted backend is required, as records can be saved to local files, warehouses, or observability systems.
AUDR emits records asynchronously and out of band, avoiding added synchronous latency on the request path unless optional pre-flight budget gating is enabled.
Adapters capture execution identifiers, token usage, tool calls, and operational timings, and they do not record prompt inputs or generated outputs.
