See what your agent actually did
Every model call, tool step, and retrieval in one trace, with tokens, cost, latency, and errors attached. For agents and any LLM app. OpenTelemetry-native, first trace in five minutes.
npm i @telemetry-dev/ai-sdkWorks with your stack
What ran, in what order, and where it broke
Follow a request from first prompt to final token. Every model call, tool call, and retrieval is a span, with inputs, outputs, timing, and errors in one waterfall.
Full span waterfall
See the whole execution tree of chains, tools, retrievals, and model calls, with timing for every step.
Inputs and outputs, captured
Open any span to read the exact input and output, with token counts and finish reason attached.
Multi-turn, grouped
Session IDs stitch a whole conversation into one trace, so multi-turn agents are easy to follow.
Cost accounting, down to the span
Computed server-side from real token usage and refreshed model pricing: input, output, cached, and reasoning tokens. No estimates. No client-side guessing.
Built for the 2 a.m. “why is this slow”
One slow request or a month of spend: the traces and metrics to answer it, fast.
Volume, cost, p95, errors
The four numbers you actually check, split by model, provider, and environment.
Failures surface as they happen
Failed calls, timeouts, and cost spikes show up as traces stream in. No dashboards to babysit.
Success rate
98%chat.completion · gpt-4o · 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
Flip capture on per environment to store exact inputs and outputs, and keep it off where compliance says no.
One bad trace in a million? Found.
Filter by model, provider, environment, status, user, or session, and open it in seconds.
Built on OpenTelemetry
Use our LangChain and Vercel AI SDK packages, or point any OpenTelemetry exporter at our endpoint. Standard OTLP in, no lock-in.
Any framework. Any language. One endpoint.
Anything that speaks OTLP can send traces. We normalize them into one model of spans, tokens, and cost.
Vercel AI SDK
TypeScript · generateText, streamText
LangChain
Python · agents, RAG, tools
OpenTelemetry
Any language · OTLP/HTTP
Tokens (30d)
4.2M
Cost (30d)
$1.2K
p95 latency
1.8s
By model
By provider
By environment
Your first trace is five minutes away
Install one package, add an API key, send a request. That's the whole setup.