Set up LLM tracing with OpenTelemetry

Connect your AI application to telemetry.dev to inspect model calls and agent steps. Use a provider integration, the TypeScript or Python SDK, or an existing OTLP/HTTP exporter.

How to send your first trace

Make a project API key from your project's API keys page.
Configure the integration for your SDK or framework with that key.
Run an instrumented request, then flush the exporter before your process exits.
Open the trace in your project to inspect its spans and reported usage.

API reference

The OpenAPI specification describes POST /v1/traces, POST /v1/logs, and POST /v1/metrics on https://ingest.telemetry.dev. The API accepts OTLP/HTTP with protobuf or JSON payloads. The ingest health endpoint uses GET /health.

Authentication

Send an Authorization: Bearer header with a project API key (td_live_...). Keep the key on your server. Keys apply to one environment or to all environments in a project. For an environment-scoped key, set deployment.environment.name to the matching environment.

Choose an integration guide

Use the guide that matches the code you want to trace:

OpenAI or Anthropic Claude: instrument the provider SDK in TypeScript or Python.
Vercel AI SDK: add the integration for your AI SDK major version.
LangChain and LangGraph: connect OpenTelemetry instrumentation or trace the provider calls.
TypeScript SDK or Python SDK: add spans around application code.
OpenTelemetry: configure an OTLP/HTTP exporter or connect your existing trace provider.

The full integration directory includes more providers, frameworks, and coding agents.

What will appear in a trace?

A trace shows the spans your instrumentation emits, with their parent relationships, timing, and errors. Model names and token counts depend on the integration and provider. Cost comes from reported usage and matching model pricing, unless the span supplies an explicit cost.

What happens to prompts and responses?

Content capture is on by default. The per-environment switch drops recognized prompt, response, and tool-payload fields before storage, not before transmission. Logs, errors, and custom attributes can still contain sensitive text. Redaction patterns cannot catch all sensitive data. To prevent transmission, remove sensitive content in your instrumentation before export.

Versioning and deprecation policy

The ingest API is versioned by URL path (/v1). Breaking changes ship under a new /vN prefix while the previous version keeps working; this page and the OpenAPI document announce each deprecation, and deprecated versions answer with a Sunset header once a removal date is set.

For AI agents

Start from llms.txt for a machine-readable site overview, use pricing.md for plan limits, and use sitemap.xml for the full list of pages.