End-to-end call chains — every MCP tool call, LLM call, and retrieval step linked by a single trace ID
MCP Tool Call (purple)
The agent retrieved context before calling the LLM. Think of it as a search step — it grabs relevant knowledge from your MCP server so the LLM gives a better answer.
LLM Call (blue)
The actual language model request — your message plus the MCP context sent to the LLM, response returned. This is where tokens are consumed and cost is incurred.
Knowledge / Web (green/orange)
Future span types for knowledge base lookups and web search — same idea, different source.
What is a Trace?
One user message = one trace. A trace contains multiple spans — the individual steps the agent took to respond. A healthy trace looks like: MCP retrieve → LLM call.
Reading the waterfall bar
The coloured bar shows when each span ran and how long it took relative to the total trace time. Wider = slower. The MCP bar should be left of the LLM bar.
Clicking a span
Click any span header to expand it and see the raw metadata — tool arguments, model name, HTTP status, exact cost.
Loading…