Skip to main content

Logs, metrics & tools

Availability

Open core · self-host + all Splyntra Cloud plans

Beyond individual traces, three screens give you the aggregate and correlated views of your agent fleet.

Logs​

The Logs screen shows structured, trace-correlated logs: every log line links back to the run and span that emitted it, so you can pivot from a log to its trace and vice versa. Logs are severity-filtered (debug / info / warn / error / fatal) and redacted by default — the same redaction that runs on trace content, so secrets and PII don't leak into your logs.

Emit them with the SDK log helpers, which attach trace context automatically:

from splyntra import log
log.info("Refund processed", {"amount": 80, "order_id": "..."})

Metrics​

The Metrics screen plots the fleet's health over time:

  • Latency — per-run and per-span duration trends
  • Throughput — run and call volume
  • Error rate — failed runs and calls
  • Spend — token cost over time

Scope by project and environment to compare, say, staging against production.

Tools & Retrieval​

The Tools & Retrieval screen focuses on the non-LLM work agents do:

  • Tool calls — invocations captured as tool_call spans
  • RAG retrieval — retrieval spans from retrievers
  • Vector search — vector_search spans from vector stores

For each it surfaces latency, failures, and flagged (risk-triggering) calls, so a slow retriever or a failing tool stands out without opening individual traces. Per-server MCP breakdowns live on the MCP Servers screen.

Next steps​

  • Traces — jump from an aggregate signal into a single run.
  • Costs — turn the spend metric into per-model attribution.
  • Alerts — get notified when error rate or risk crosses a threshold.