Analytics preference

We only load Google Analytics after you opt in. You can change this choice at any time from Cookie settings.

Traces Tab

The Traces tab shows LLM execution traces.

The Traces tab shows LLM execution traces. Inspect request/response payloads, timing, tool calls, and debug Agent and LLM node behavior, plus aggregate metrics (calls, tokens, cost, latency, error rate) across any time window.

Stats Header

The page opens with a stats summary above the trace list, computed across the selected time range and other active filters:

  • KPI cards – Total Calls, Total Tokens, Total Cost (USD), Avg Latency, Error %
  • Tokens by Model – Donut chart of token volume per model
  • Cost by Model – Donut chart of USD spend per model. Models without pricing show a "Configure pricing" link to the LLM Cost Table
  • Calls Over Time – Stacked area chart of successful vs failed calls per bucket (5 min / 1 h / 6 h / 1 day, chosen automatically per range)
  • Unpriced models notice – Appears below the charts when some traced models have no pricing row, linking to the LLM Cost Table

Time Range

A single Time range selector at the top of the filter bar drives both the stats header and the trace list. Choices: Last 1 hour, Last 24 hours, Last 7 days (default), Last 30 days, All time. Changing the range resets pagination and reloads charts and list together.

Trace List

  • Paginated list of recent traces within the selected time range
  • Each trace shows: workflow, node, credential, timestamp, duration
  • Click a trace to open the detail view
  • Use Prev / Next inside the detail dialog to move between traces without closing it

Filtering

  • Time range – 1h / 24h / 7d / 30d / All; filters charts and list together
  • Credential – Filter by credential (API key) used
  • Source – Filter by workflow or execution source
  • Search – Search by model, workflow, credential, or node label
  • Refresh – Reload the current page of traces

Trace Detail

When you open a trace:

  • Steps – A top-to-bottom timeline that reads the trace as an ordered chain: System → User → Assistant → Tool → Answer. Each step is collapsible; expanding one shows a readable detail (message text, or a tool's arguments and result) followed by that step's JSON. JSON content in every event type—including MCP calls and JSON-RPC payloads—opens in an expandable tree by default, with the first level expanded. Use Tree / Raw to switch between the formatted structure and the original text. Malformed JSON remains visible as plain text. Tool steps display their duration and an MCP/Skill badge where applicable, plus pending / timeout / cancelled badges when the tool-call status is set. The Answer step can show aggregate tool_metrics badges (call counts by status and total tool duration). Traces without a message chain (for example image generation) show a minimal Request → Response timeline.
  • Request – Full request payload sent to the LLM, with Tree / Raw JSON views
  • Response – Model response, including tool calls if any, with Tree / Raw JSON views
  • Tool metrics – When present on the response, a compact summary of tool call counts (success / error / pending / timeout / cancelled) and total tool duration
  • Timing breakdown – llm_ms, tools_ms, mcp_list_ms for performance analysis
  • Tool calls – Tool name, arguments, result, status, and elapsed time
  • Skills included – Skills passed to the model in the request
  • Go to Workflow – Jump directly to the related workflow from the trace detail dialog

Copy and Export

  • Copy request or response JSON to clipboard; copying always uses the complete raw payload regardless of the active view
  • Use for debugging or sharing with support

Maintenance

  • Clear All – Delete all saved traces after confirmation
  • Pagination controls show the current visible range (for example 1-25 of 240)