Skip to content
telemetry.dev
Esc
↑↓navigate↵open⌘Jpreview
On this page

telemetry.dev documentation

Send your first AI trace, then connect the provider, framework, or coding agent you already use.

telemetry.dev records what your AI application does. Each run becomes one trace that shows every model call, tool call, and step in order, with the tokens, cost, latency, and errors of each one.

A telemetry.dev trace named support-agent with a success status: a tree of 13 spans beside a timeline, with HTTP calls that load the customer profile and session, model calls, a retrieval step with embeddings, the search_help_center tool, and a guardrail check, each with its duration.
A trace in the dashboard. Each row is a span, nested under the span that started it, with a bar for when it ran.

telemetry.dev is built on OpenTelemetry. The SDKs send standard OTLP/HTTP, and so can the OpenTelemetry setup you already run.

Get started

Quickstart

Send a real trace and find it in your project in about five minutes.

How it works

Traces, spans, environments, and what the ingest service does with your data.

Connect your application

Model provider or framework

Wrap the OpenAI, Anthropic, Gemini, or Bedrock client, or add tracing to the Vercel AI SDK, LangChain, and other frameworks.

Coding agent

Record Claude Code, Cursor, opencode, and other agents on your machine.

TypeScript SDK

Record functions, model calls, tools, and agents on Node.js, Vercel Edge, or Cloudflare Workers.

Python SDK

Record synchronous and asynchronous Python code with a decorator or context managers.

Existing OpenTelemetry

Keep your instrumentation and add telemetry.dev as an OTLP/HTTP destination.

What you can do

  • Debug a single run. Open a trace to read each model call’s input and output, each tool’s arguments and result, and the error that ended a failed run.
  • Follow a conversation. Traces that share a session ID are grouped into one session.
  • Track cost and usage. Tokens and cost are calculated per span and added up per trace, session, model, and user.
  • Keep sensitive data out. Turn off prompt and response capture per environment, and redact secrets and personal data before storage.
  • Measure quality. Run a prompt version or your code on a dataset, score the replies with LLM judges, code checks, and reviewers, and compare the results.

Platform

Traces, sessions, and logs

Search traces, read the span waterfall, and follow sessions.

Analytics

Requests, cost, p95 latency, tokens, and error rate over time.

Evals

Datasets, experiments, evaluators, scores, and review queues.

Privacy and retention

Control what is captured, redact sensitive values, and set retention.

OTLP ingest API

Endpoints, authentication, limits, and error responses.

Last updated on October 30, 2026

Was this page helpful?