llm-cost-telemetry · packages

Core types, Zod schemas, shared utilities, and configuration loaders that serve as the foundation for the `@reaatech/llm-cost-telemetry-*` ecosystem. It exports 40+ domain types (CostSpan, BudgetConfig, TelemetryContext, etc.), 35 Zod schemas for runtime validation, 25 utility functions (ID generation, time window math, cost calculation), and typed configuration loaders, with only `zod` as a runtime dependency.

A set of classes (`CostCollector`, `CostAggregator`, `BudgetManager`) for buffering, aggregating by tenant/feature/route/model, and enforcing per-tenant daily/monthly budgets with cascading alert thresholds on LLM cost telemetry spans.

A function that calculates LLM API costs across OpenAI, Anthropic, and Google models, with built-in pricing tables, cache-aware pricing, token counting, and pre-call cost estimation.

A CLI tool that generates LLM cost reports, checks budget status, and triggers exports to observability platforms (CloudWatch, Cloud Monitoring, Phoenix/Loki) from JSON cost span files or stdin.

Exporters for pushing LLM cost telemetry to AWS CloudWatch (standard and EMF), GCP Cloud Monitoring, and Grafana Loki/Phoenix, provided as configurable class instances (`CloudWatchExporter`, `CloudMonitoringExporter`, `PhoenixExporter`) that extend a shared `BaseExporter` with batching and retry logic. Requires the respective cloud SDKs (`@aws-sdk/client-cloudwatch`, `@google-cloud/monitoring`) as peer dependencies.

An MCP server that exposes three layers of LLM cost telemetry tools — atomic span recording, multi-dimensional aggregation, and budget enforcement — consumable by MCP clients like Claude Desktop and Cursor.

A set of OpenTelemetry tracing, metrics, and Pino-based logging managers specifically for tracking LLM API costs, providing classes like `TracingManager`, `MetricsManager`, and a `getLogger()` singleton that automatically record token usage, cost amounts, and PII-redacted logs with Gen AI semantic conventions.

Provides `wrapOpenAI`, `wrapAnthropic`, and `wrapGoogleGenerativeAI` functions that wrap their respective official SDK clients to automatically emit `CostSpan` objects with token usage, timing, and telemetry context from every API call.