Introduction

Public packages for product LLM attribution and coding-agent spend collection.

trAIce connects AI usage to the business dimensions that explain it: customer, user, feature, workflow, agent, and team. The public repository provides product-runtime SDKs, a direct HTTP contract, and local coding-agent collectors.

Choose a data path

Product usage and internal AI-tool usage are separate data streams. Choose the path that matches who initiated the spend.

Path	Use it for	Destination	Start here
Product SDKs	LLM calls made by your application for customers or end users	`/api/v1/events`	Product SDKs
Internal Spend collector	Claude Code and Codex usage by employees and teams	`/api/v1/internal-usage`	Collector overview
Ask trAIce	Spend questions and explicitly confirmed Team actions	`/api/v1/ask` or MCP	Ask trAIce

Do not send employee coding-agent usage as a product event. Do not use customer identifiers as employee identifiers. Keeping the two paths separate preserves clear allocation and reporting.

Public packages

Package	Purpose	Registry	Source
`@traice/sdk`	TypeScript and Node.js product-runtime metering, adapters, analytics, and request guardrails	npm	packages/sdk
`traice-sdk`	Python product-runtime collection for OpenAI, Anthropic, LangChain, and LangGraph	PyPI	packages/python
`@traice/collector`	Local Claude Code and Codex telemetry collection	npm	packages/collector
`@traice/protocol`	Shared public event types, validation, normalization, and redaction helpers	npm	packages/protocol

Product attribution model

Every product integration can attach the same dimensions:

tenantId identifies the customer or account that pays you.
userId identifies the end user who initiated the call.
feature identifies the product capability or request path.
workflowId, runId, and stepId identify multi-step executions.
agentId and toolName identify agent and tool activity.
outcome and retryCount explain result quality and retry behavior.
metadata carries additional structured context that is safe to send.

tenantId is the primary dimension for customer-level AI contribution margin. Use the same stable customer identifier that your billing or revenue data uses.

Security and privacy

Store the trAIce API key in your secret manager and provide it through TRAICE_API_KEY. Provider keys remain in your infrastructure and are never sent to trAIce.

Product attribution does not require prompts or model outputs. The Python SDK never sends either. The TypeScript SDK sends them only when you explicitly pass prompt or output and your workspace is configured to accept samples. Collectors keep prompt capture off by default.

Next steps

Create a workspace key and send a first event.
Choose TypeScript and Node.js, Python, or HTTP and cURL.
Use the API reference for signatures, event contracts, and source links.