Show HN: Observability for Coding Agents and LLM Applications

ephraimduncan1 pts0 comments

Telemetry — Observability for AI and LLM apps

Stop guessing what your AI is doing<br>Every model call, tool step, and retrieval in one trace — tokens, cost, latency, and errors attached. Built on OpenTelemetry: any framework, any language, no lock-in. First trace in five minutes.<br>Start tracing freeCopy agent prompt<br>Free plan: 10k spans a month. No credit card.

When something breaks, you'll know why

One slow request or a month of spend: the traces and metrics to answer it, fast.

Volume, cost, p95, errors<br>The four numbers you actually check, split by model, provider, and environment.

Total cost (30d)<br>$1,248<br>↑ 12% vs prev 30d

p95 latency<br>1.8s<br>↓ 240ms vs last week

Error rate<br>0.6%<br>↓ 0.3% vs last week

Failures surface as they happen<br>Failed calls and timeouts stream in as they happen, with error type, status, and the full trace attached.

Success rate<br>98%

chat.completion · gpt-5.5 · 1,204 tok · $0.004

retrieval · pinecone · 8 docs · 120ms

tool · get_invoice · running

User message

LLM call · 820ms

Response streamed

Inputs and outputs, on your terms<br>Flip capture on per environment to store exact inputs and outputs. Secrets are redacted at ingest, and you can add your own patterns.

gpt-5.5chat.completion

200 OK<br>system<br>You are a support assistant. Reply in two short lines.

user<br>Summarize this support thread.

assistant<br>Payment failed at checkout — the customer's card was declined. They want a retry link.

1,204 tokens · $0.004 · 820ms

One bad trace in a million? Found.<br>Search by name, filter by status, environment, framework, or session, and open the full trace in seconds.

gpt-5.5productionerror<br>Matching traces<br>312

Error rate<br>5.8%

Send traces from anywhere<br>Drop in our SDK, or point any OpenTelemetry exporter at our endpoint. Any framework, any language, no lock-in.

Integrations<br>Any framework. Any language. One endpoint.

However you added AI to your app, send the traces here. We normalize them into one model of spans, tokens, and cost.

Drop-in SDKs<br>TypeScript and Python SDKs: wrap a function in observe() and it becomes a span, with inputs, outputs, tokens, and errors captured. One install, one API key.<br>Standard OTLP ingest<br>No SDK required. Point any OpenTelemetry exporter at our endpoint and traces arrive normalized.<br>Token and cost trends<br>Track input, output, cached, and reasoning tokens over time, with cost computed per call from live pricing for 4,700+ models.<br>Group by anything<br>Break cost and latency down by model, provider, environment, or user. No setup.

Vercel AI SDK<br>TypeScript · generateText, streamText

Connected<br>LangChain<br>Python · agents, RAG, tools

Connected<br>OpenTelemetry<br>Any language · OTLP/HTTP

Connect

Tokens (30d)<br>4.2M

Cost (30d)<br>$1.2K

p95 latency<br>1.8s

By model<br>gpt-5.5<br>claude-sonnet-4-6

By provider<br>OpenAI<br>Anthropic

By environment<br>production<br>staging

Frequently Asked Questions<br>What is telemetry.dev?+telemetry.dev is an observability platform for AI and LLM apps. It captures every model call, tool step, and retrieval as OpenTelemetry traces, then attaches tokens, cost, latency, and errors — so you can debug failures and track spend across models, providers, and environments.<br>How do I start sending traces?+Drop in our SDK, or point any OpenTelemetry exporter at our endpoint. Add your project API key and your first trace shows up in about five minutes.<br>Which frameworks and languages are supported?+TypeScript and Python have first-party SDKs, and the Vercel AI SDK has a drop-in integration. Anything else that emits standard OTLP over HTTP — LangChain, OpenLLMetry-instrumented apps, or your own OpenTelemetry setup — works through the same endpoint. More drop-in integrations are on the way.<br>How is cost calculated?+Cost is computed server-side from each call's real token usage — input, output, cached, and reasoning — against current per-model pricing for 4,700+ models. There are no client-side estimates to keep in sync, and you can override cost per span when you need to.<br>Is the free plan actually free?+Yes, with no credit card. You get 10k spans a month, 7-day retention, one project, and two seats, free for as long as you want. Upgrade only when your volume outgrows it.<br>Can I keep prompts and completions out of storage?+Yes. Input and output capture is a per-environment switch, so you can store full prompts in development and none in production. Secrets like API keys are redacted at ingest, and you can add custom redaction patterns per project.

Your first trace is five minutes away<br>Install one package, add an API key, send a request. That's the whole setup.<br>Start tracing freeTalk to an engineer<br>Free plan includes 10k spans a month. No credit card required.

cost model trace tokens opentelemetry traces

Related Articles