Get started
Your first trace
Open the first trace the SDK sent, read the model call inside it, and see what the tokens, cost and tabs of a trace mean.
A trace is one request through your application, or one agent turn, recorded as a tree of spans. Your first trace is the request you made at the end of Install the SDK.
Find it
On the quickstart page, click View trace next to "First trace received". Otherwise open Traces and click the row.
The trace list shows time, operation, service, status and latency for the selected Environment. A Flags column appears when a listed trace is flagged: error, empty reply, truncated, tool loop or model swapped. A Score column appears once a listed trace has been evaluated.
Read it
The header shows the trace id, the wall time and the span count. Tokens and cost appear when the trace has a model call, and subagents and errors when there are any. Session and User show the ids your code set.
| Tab | What it shows |
|---|---|
| Conversation | An agent run as a conversation. Only agent runs have this tab, and it opens first when the trace has 500 spans or fewer. |
| Waterfall | The span tree in time. The legend has a chip for each span type present, with its count, and a Critical path chip. Click a chip to filter the tree. |
| Logs | The log lines the SDK captured, in order. |
| Evaluation | The scores, once something has evaluated the trace. |
| Metadata | Service, Environment, start time, SDK, and counts of model calls, tool calls and agents, then up to 40 attributes of the root span. |
Read the model call
Click the llm_call span. Under the model and provider, the panel shows:
- the request: the sampling parameters as sent and the messages as a chat transcript
- the completion, and a thinking block when the model produced one
- tokens in and out, with reasoning, cache-read and cache-write tokens listed separately, and the cost
- time to first token, when the call streamed
- the tool calls the model invoked, with their arguments and results
- the raw request and response bodies
Cache-read and cache-write tokens are subsets of the input tokens, so they are shown beside the total, not added to it.
If the request text is missing
A service started with capture_content=False still records tokens, latency, cost, tool names and the span tree. In place of the request, the panel says "No request text arrived for this call" and names the setting that turns capture on: capture_content=True. For a coding agent it names evalkit coding install --vendor <vendor> --capture full and tells you to restart the tool. See Configuration and privacy.
What changes on Home
Once the Environment has a trace, Home shows its health over a time range you pick. The default is the last 24 hours, compared with the 24 hours before. Home then has:
- tiles for requests, error rate, p95 latency and estimated cost, each with a trend and its change against the previous period
- a chart of spans, error spans and p95 latency over time: click a bar to open its traces, drag to zoom and double-click to reset
- model-call figures (calls, tokens, LLM p95, average time to first token and calls by model), when the window has model calls
- the recent traces, runs and alerts
What the checklist does next
Until you finish or skip it, Home also shows the Getting started card. Its five steps are Environment key, Install SDK, Receive first trace, Add an evaluator and Run first evaluation. The app checks them against the project's data on every load.
Install SDK and Receive first trace are done now. Add an evaluator was done from the start, because every organization begins with managed evaluators and collections. Run first evaluation continues in Your first evaluation.
Skip for now hides the card for this project, in this browser only.