telemetry.dev

telemetry.dev is an observability platform built specifically for AI and LLM applications. It captures every model call, tool step, and retrieval as OpenTelemetry traces, attaching token counts, cost, latency, and error details to each span. The platform is designed to get developers from zero to their first trace in about five minutes, with no credit card required to start.

What It Is

telemetry.dev sits in the observability layer of an AI application stack, giving developers a unified view of what their models, tools, and retrieval systems are doing at runtime. Rather than building on a proprietary protocol, it is built on OpenTelemetry — meaning any framework or language that can emit standard OTLP over HTTP can send data to it without vendor lock-in. The platform normalizes all incoming spans into a single model of spans, tokens, and cost, then surfaces the four metrics the team describes as the ones developers actually check: volume, cost, p95 latency, and error rate.

How Tracing Works

The core unit of data is the span — one step in a trace, such as a model call, a tool call, or a retrieval. An agent run with five steps produces five spans. Developers can instrument their apps in two ways:

• Drop-in SDKs: TypeScript (@telemetry-dev/sdk on npm) and Python (telemetry-dev on PyPI) SDKs let developers wrap a function in observe() to turn it into a span with inputs, outputs, tokens, and errors captured automatically.
• Standard OTLP ingest: Any OpenTelemetry exporter pointed at the ingest endpoint (ingest.telemetry.dev) sends traces without requiring the SDK. LangChain, OpenLLMetry-instrumented apps, and custom OpenTelemetry setups all work through the same endpoint.

Cost is computed server-side from each call's real token usage — input, output, cached, and reasoning tokens — against current per-model pricing for what the site describes as 4,700+ models, with no client-side estimates to maintain.

Integrations and Framework Support

The platform lists first-party or drop-in support for a wide range of model providers and frameworks, including OpenAI, Anthropic, Google Gemini, Mistral AI, Cohere, Meta Llama, Hugging Face, Groq, Ollama, xAI, DeepSeek, Perplexity, LangChain, OpenRouter, Together AI, and the Vercel AI SDK. The ingest API is versioned by URL path (/v1) with a stated deprecation policy: breaking changes ship under a new /vN prefix while the previous version continues to work.

Privacy and Data Controls

Input and output capture is a per-environment toggle, so teams can store full prompts in development while capturing nothing in production. The platform redacts secrets like API keys at ingest and allows custom redaction patterns per project. This makes it practical for teams with compliance requirements around prompt and completion storage.

Audience and Deployment Model

telemetry.dev targets developers and engineering teams shipping AI features in production. The platform is delivered as a cloud-hosted SaaS — there is no self-hosted option described in the available sources. The free plan is positioned for side projects or single services, while higher tiers are aimed at teams running AI across multiple services and environments. The API reference is described by an OpenAPI specification covering /v1/traces, /v1/logs, /v1/metrics, and /health.