Dashboard reference

Explore Metergraph walks through the dashboard as one team uses it. This page is the reference for the details that tour leaves out.

Spend by model

Model and provider breakdowns show where spend went and which calls could not be priced. For catalog coverage and the exact pricing rules, see Model catalog and pricing.

Traces

Traces are the evidence underneath the summaries. Open them from a workload or use the trace explorer when you need to search across workloads. A trace shows the model calls, timing, token use, cost, errors, and captured request and response content that make up one unit of work. Prompts and responses appear only where they were captured; see Content and privacy.

Filters

The explorer filters by time window, environment, workload, provider, model, status, captured content and exact metadata pairs. A metadata filter keeps calls whose metadata has the key with exactly that value. Matching is case-sensitive, and up to five pairs combine. Environment, provider, model and metadata filters select calls, so a listed trace’s figures cover its matching calls. Search matches a trace id prefix or a trace name.

Spread samples

Choose Spread sample instead of Newest first to see traces drawn evenly across the window rather than the most recent page. Metergraph splits the window into ten equal time slices and takes the matching call nearest the middle of each, one trace per call. A sample therefore covers the whole window, and the same window and filters always draw the same traces. Opening a sampled trace shows the call that matched. A sample holds at most ten traces, so it is not every matching trace, and search does not narrow it.

Compare groups

Compare groups draws one spread sample for each of two to four groups over the same window and environments. A group is a set of filters: workloads, providers, models, metadata pairs, failed calls, captured content, and a range of cost or latency per call. Read side by side shows the trace each group drew from one time slice, with its prompt and response, and steps through the window slice by slice. The page’s address keeps the groups, so a comparison can be shared with anyone in the workspace.

You can also start a comparison from Compare samples by model on a workload’s Models table, or from Compare samples on Performance when two or more rows are selected. An alert’s Compare traces view does the same for one alert: traces from the alert window beside traces from its normal baseline.

Pricing

Pricing shows the models and channels Metergraph can price. Use it to confirm coverage or understand why a call is marked unpriced. Pricing is applied when a call is ingested and is not recalculated later.

Ingest keys

Open Keys to create or disable credentials for sending traces. Ingest keys can write telemetry but cannot read it back. Coding-agent connections are separate, and you choose their access when you create one. An owner’s new connections default to Debug and can instead be Metadata or Replay. A member can create Metadata connections only. Metadata reads content-free telemetry (agent:metadata), Debug adds retained trace and report reads (agent:read), and Replay adds agent:replay on top of Debug. See Manage your workspace.

Batch API

Batch API reports platform-wide completion times from one synthetic, single-request probe per provider every four hours. The probes are never your traffic and do not measure your workloads.

The Direct-fallback deadline control runs from 3m to 20m. It is the point when a caller would stop waiting for the batch and send a normal request, not a maximum wait applied to the probes. Metergraph uses the observed completion distribution to estimate cost at that deadline. The page is hidden on BYOC deployments.

Settings

Settings holds workspace membership, defaults, and analysis configuration. Availability depends on your deployment profile and role.