llm-cost-telemetry · packages

Provides shared TypeScript types, Zod schemas, and utility functions for modeling, validating, and calculating LLM cost telemetry data. It serves as the foundational library for the broader telemetry ecosystem, requiring only Zod as a runtime dependency.

Aggregates LLM cost telemetry across multiple dimensions and enforces per-tenant budgets using a set of classes for buffered span collection, stateful data grouping, and threshold monitoring. It provides the `CostCollector`, `CostAggregator`, and `BudgetManager` classes to manage, track, and alert on usage costs in real-time.

Calculates LLM API costs and provides pre-call budget estimations using a library of built-in pricing tiers for OpenAI, Anthropic, and Google models. It exposes a set of utility functions for calculating token-based costs, comparing model pricing, and applying custom pricing overrides.

A CLI tool that reads LLM cost span data from

Provides exporter classes (CloudWatchExporter

Exposes LLM cost tracking, aggregation, and budget enforcement as a set of Model Context Protocol (MCP) tools. It provides a factory function that returns an MCP server instance, designed to be integrated with MCP-compliant clients like Claude Desktop or Cursor.

Instruments LLM applications with OpenTelemetry tracing, cost-tracking metrics, and PII-redacting Pino logging. It provides `TracingManager` and `MetricsManager` classes to handle OTLP exports and Gen AI semantic conventions.

Wraps official OpenAI, Anthropic, and Google Generative AI SDK clients to automatically capture token usage, latency, and custom metadata for every API call. It provides wrapper functions that return augmented client instances, requiring the respective provider SDKs as peer dependencies.