LLM observability for AI agents
Follow model calls, tool steps, and retrievals in one trace. Find failures, inspect latency, and track token usage and cost. Connect with our SDKs or your existing OpenTelemetry exporter.
Free plan: 30,000 ingestion units a month, one per span, log record, or metric data point. No credit card required. Read the quickstart.
Support Agent
Find the slow step. Inspect the failed call.
Pinpoint a tool timeout, inspect a model response, and see where your budget goes.
Find the slow step
Follow model calls, tool steps, and retrievals in one trace, with the duration of each step.
- track_shipment
- 2,418ms
- 49%
Keep cost and reliability in view
Compare spend, latency, and errors by model, provider, and environment.
- 812
- 1.8s
- 2.6M
Find the trace you need
Search by name, or filter by status, environment, framework, or session.
- support-agent
- 9
- $0.0124
See where MCP calls fail
Spans, failures, and p95 duration by method and tool, from the client and the server.
See MCP analytics| Tool | Spans | p95 | Failed |
|---|---|---|---|
| create_refund | 2,330 | 2.4s | 18.0% |
| get_invoice | 5,912 | 640ms | 1.1% |
| check_inventory | 4,208 | 310ms | 0.2% |
| cancel_order | 1,077 | 880ms | 0.6% |
Hear about it before your users do
Email alerts for cost, budgets, agent loops, and model or tool errors.
See alerts- Tool errorsAlerting
- Monthly budgetAll clear
- Agent loopAll clear
- Missing pricesNeeds more data
Control what gets stored
Turn off prompt capture for an environment before storage. Redaction does not catch all sensitive data.
Connect your application
Use our TypeScript or Python SDK, or your existing OTLP/HTTP exporter.
Find your setup guide- Any language
Start building
Give the setup to your coding agent, or open the guide for the stack that you use.
Send your first trace
One prompt for your coding agent
- 01Finds your stack
- 02Adds the right integration
- 03Sends a test trace
- 04Lists the variables to set
Model providers
8You call a provider SDK.
Frameworks
7Runs your calls and tools.
SDKs and OpenTelemetry
3Your spans, or your exporter.
Not on the list? Any exporter that sends OTLP/HTTP works.
OTLP ingest APIAlready sending traces? Let your AI agent investigate errors, latency, and cost through telemetry.dev’s read-only MCP server. Connect an AI agent.
Frequently asked questions
Still have a question? Contact us
What is telemetry.dev?
telemetry.dev is an LLM observability platform for AI applications and agents, built on OpenTelemetry. A trace groups the instrumented model calls, tool steps, and retrievals in a request or agent run. Reported token usage, cost, latency, and errors help you debug failures and monitor spend.
How do I start sending traces?
Make a project API key, then configure a telemetry.dev integration or an OTLP/HTTP exporter. Run an instrumented request and open its trace in your project. The integration directory links to setup instructions for each SDK, provider, and framework.
Which frameworks and languages are supported?
TypeScript and Python have first-party SDKs, with integrations for providers and frameworks such as OpenAI, Anthropic, and Vercel AI SDK. The Claude SDK and the Claude API can also send their native OpenTelemetry spans to telemetry.dev. Other languages can send standard OTLP over HTTP. Trace detail depends on what your instrumentation emits, so each integration guide describes its coverage.
How is cost calculated?
telemetry.dev calculates cost at ingest from reported token usage and its model-pricing catalog. Cache-read and cache-write tokens use separate rates when available. Reasoning tokens are not charged separately. An explicit span cost takes precedence. Without matching model pricing or token usage, computed cost is unavailable.
Is the free plan actually free?
Yes. The Free plan costs $0 and requires no credit card. It includes 30,000 ingestion units per month, 14-day retention, one project, and two seats. Each normalized span, OTLP log record, or metric data point uses one ingestion unit.
Can I keep prompts and completions out of storage?
Yes, for recognized prompt, response, and tool-payload fields. Capture is on by default, and you can disable it per environment before storage. This does not stop transmission to ingest or remove arbitrary content from logs, errors, or custom attributes. Built-in and project-specific redaction patterns apply, but pattern matching cannot catch all sensitive data. To prevent transmission, remove sensitive content in your instrumentation before export.
Trace your next AI request
Choose an integration, add your project API key, and run an instrumented request.
Start free
30,000 ingestion units a month, one per span, log record, or metric data point. No credit card required.
npm install @telemetry-dev/sdkTalk to an engineer
Our team shows you tracing, cost tracking, and privacy controls for the frameworks and model providers that you use.