Skip to content

Traces

The Traces tab shows every API call routed through your proxy endpoints. Each trace captures the full round-trip: what you sent, what Claude returned, how many tokens were used, and what it cost.

Browsing traces in the dashboard

Navigate to Traces in the dashboard to see all API calls, newest first.

Each row shows:

ColumnDescription
TimestampWhen the request was made (click to view detail)
ModelThe model that handled the request
TokensInput / output token counts
CostCalculated cost based on model pricing
DurationRound-trip time to Anthropic
StatusHTTP status code (200 = success)

Use the filter bar at the top to narrow results:

  • Model — filter by model name (e.g. “opus”, “sonnet”, “haiku”)
  • From / To — date range filter

Click any trace to see the full request and response.

The token card shows:

  • Input tokens — tokens in the prompt
  • Output tokens — tokens in the response
  • Cache created — tokens used for prompt cache creation
  • Cache read — tokens read from prompt cache
  • Total cost — calculated from model pricing

The left panel parses your request JSON and shows:

  • Model and parameters — model name, max_tokens, thinking config
  • System prompt — collapsible, so long system prompts don’t dominate the view
  • Tools — list of tool definitions sent to the model
  • Messages — the conversation, with role badges (user/assistant)

The right panel shows what Claude returned:

  • Streaming responses — SSE events are parsed and reconstructed into the final message, showing text output, thinking blocks, and tool use calls
  • Non-streaming responses — JSON is formatted and displayed
  • Errors — red banner with the error body for failed requests (4xx/5xx)
  • Raw toggle — click “Show raw SSE” to see the original SSE event stream

Filter traces by status code. Click a failed trace to see the error body in the response panel. The request panel shows exactly what was sent, so you can reproduce the issue.

Look at the token breakdown. High cache_read tokens with low cache_creation tokens means your prompt cache is warm — subsequent requests are cheaper. If you see high cache_creation on every request, your prompts may be changing too frequently.

Create separate proxy endpoints for each project or environment (e.g. “production”, “staging”, “dev”). Each endpoint’s usage is tracked independently and visible in the traces filter.