Cost on every call
Spancache prices each span from its tokens — cache-aware — so you get accurate per-call, per-trace, per-model spend even when the source reports no dollars (like Claude Code on a subscription). See Cost analytics.
Spancache
See what every agent run costs, why it was slow, and why it failed — on a store you can keep forever.
Observability for AI agents
No SDK — OTLP-native
Point any OpenTelemetry exporter at one endpoint. Works with Claude Code, the OpenAI/Anthropic SDKs, LangChain, LlamaIndex and the Vercel AI SDK. Nothing proprietary to install or review. See Sending traces.
Guardrails & anomalies
One-click detection templates for agent failure modes — runaway cost, tool-error loops, retry storms, truncation — plus cost-outlier traces and novel-behavior anomalies. See The console.
Keep every trace
Losslessly compressed and never sampled, so you keep complete history at a fraction of typical observability cost — not a 7-day window. See Cost analytics.
Metadata-first governance
Stores cost, tokens, latency and outcomes by default — no prompt or response content. Content is an explicit opt-in, and anything captured is scrubbed for secrets at ingest. See Security.
Query everything
Search spans and aggregate them — kind=span | stats sum(cost) by model. The Traces and Cost
views run on the same engine; no sampling, no separate index. See Querying.