Developer + API

How do you observe Plugsky usage, latency and cost in production?

Plugsky logs every request with timestamp, model, tokens, latency, status, key ID, project, region, request ID and your optional user tag. The dashboard turns that into usage and analytics views, and logs export to Datadog, Splunk, Grafana, New Relic, OpenTelemetry or your SIEM. Because plans are flat-rate, the metrics that matter most are latency, error rate, model distribution and per-key attribution — not token billing.

Key facts

Logged fieldsTimestamp, model, tokens, latency, status, key ID, project, region, request ID, user tag
Export targetsDatadog, Splunk, Grafana, New Relic, OpenTelemetry, SIEM
Dashboard viewsUsage and analytics pages with per-model and per-key breakdowns
Usage APIGET /v1/plugsky/usage is documented as coming soon
Token countsusage.prompt_tokens and usage.completion_tokens in every response
Billing modelFlat monthly plans with fair-use usage; token counts do not drive billing
TracingRequest ID correlates a client call with platform logs
Product statusLive (dashboard and logs); usage API coming soon

TL;DR

  • Every request is logged with enough fields to attribute and debug it.
  • Export logs to your existing APM or SIEM instead of building a new pane of glass.
  • Track latency percentiles, error rates and model distribution — not just volume.
  • Use per-key attribution to find the workload behind a spike.
  • The usage API is coming soon; dashboard analytics and log export work today.

How it works, step by step

  1. Standardise a request ID and user tag on every client call for correlation.
  2. Export Plugsky logs to your APM or SIEM and confirm fields arrive intact.
  3. Build dashboards for p50/p95 latency, error rate and model distribution.
  4. Break usage down by key and project to attribute traffic to services.
  5. Alert on 429 and 5xx rates and on time-to-first-token regressions.
  6. Review token usage per response as a prompt-efficiency signal, not a billing metric.
  7. Revisit thresholds quarterly as workloads and models change.
1Standardise arequest ID and usertag on every client2Export Plugsky logsto your APM or SIEMand confirm fields3Build dashboardsfor p50/p95latency, error rate4Break usage down bykey and project toattribute traffic5Alert on 429 and5xx rates and ontime-to-first-token6Review token usageper response as aprompt-efficiency

Try it yourself

Open the LLM token calculator →

What Plugsky logs, and why each field matters

  • Timestamp, request ID, status: the spine of any trace — join client logs to platform logs with the request ID.
  • Model and region: failover and routing can change which model answered; without this field, performance shifts are unexplainable.
  • Tokens and latency: prompt and completion tokens support prompt-efficiency work; latency feeds the user-experience SLO.
  • Key ID, project and user tag: attribution. Per-key breakdowns turn "the API is slow" into "the batch worker on key 3 is saturating embeddings".

Logs export to Datadog, Splunk, Grafana, New Relic, OpenTelemetry and general SIEM pipelines, so observability stays in the tools your on-call already uses.

The metrics that actually predict trouble

Because self-serve plans are flat-rate, token spend is not the alarm to build. Track time-to-first-token for streaming, end-to-end latency percentiles, 429 and 5xx rates, and the distribution of models answering requests. A rising p95 with a stable p50 usually means queuing or retry storms; a model-distribution shift after a config change usually explains a quality regression. Pair platform metrics with an application-level quality signal — thumbs, eval scores or task success — because everything can look healthy while answers get worse.

Per-key attribution and governance

Scoped keys are the unit of attribution. Give each service, environment and agent fleet its own key, then build a dashboard per key: requests, error rate, latency, models used. This makes capacity conversations concrete — you can see whether chat, embeddings or an agent loop is driving load — and it bounds incidents, because revoking one key is a contained action. For enterprise deployments, audit logs capture key creation, rotation, scope changes and deletion with actor, timestamp, IP and request-body hash, exportable to your SIEM.

Closing the loop

Observability only pays off when it changes behaviour. Wire alerts to on-call with clear runbooks: 429s mean back off and consider a tier change; sustained 5xx means check the status page before code; latency regressions trigger a model or prompt review. Keep a monthly review of per-key usage, retired services and stale keys. The GET /v1/plugsky/usage endpoint is coming soon for programmatic reporting; until then, the dashboard plus log export covers production needs. See /docs for the current field list and export setup.

Honest comparison

CapabilityPlugskyLogging only in your appThird-party LLM gateway
Request loggingPlatform-side with model, tokens, latency, regionOnly what you remember to logVaries by vendor
AttributionKey ID, project and user tagCustomUsually key-based
ExportsDatadog, Splunk, Grafana, New Relic, OTel, SIEMYour pipelineVendor integrations
DashboardUsage and analytics viewsBuild your ownVendor dashboards
Failover visibilityModel field shows who answeredGuessworkVaries
Usage APIComing soonNot applicableVaries

Frequently asked questions

What does Plugsky log per request?

Timestamp, model, token counts, latency, status, key ID, project ID, region, request ID and an optional user tag — enough to attribute and debug traffic without application-side instrumentation.

Can I export logs to my own tools?

Yes. Logs export to Datadog, Splunk, Grafana, New Relic, OpenTelemetry and SIEM pipelines, so you can keep one observability stack.

Are token counts used for billing?

Not on self-serve plans — those are flat monthly with fair-use usage. Token counts are returned for observability and prompt-efficiency work.

Is there a usage API?

GET /v1/plugsky/usage is documented as coming soon and will return per-key, per-model, per-day usage. The dashboard and log exports cover reporting today.

How do I attribute traffic to a service?

Use a separate scoped key per service and environment, then group by key ID in exports or the dashboard. Per-key views make spikes easy to trace.

What should alert on-call?

429 and 5xx rates, latency percentiles (especially time-to-first-token for streaming), and abrupt model-distribution changes after config updates.

Does export include request and response bodies?

The documented log fields cover metadata and usage rather than full payloads. If you need prompt-level tracing, log it on your side and correlate by request ID.

How long are logs retained?

Retention details are documented in the platform docs and can be extended by enterprise arrangements. Confirm current terms for your compliance requirements.