Dashboard

Platform status#

The first thing on the page is a single sentence: Everything is running, or the number of things that need attention with a chip for each one. Every status behind it was already being recorded somewhere — by the SaaS sync, the warehouse connection test, the swarm scheduler, the pipeline run, the data monitor — and each chip links to the page that can fix it.

The four figures#

Below the band are the numbers that change a decision, each measured against something rather than standing alone.

Runs (24h)

Model calls in the last 24 hours. Shown as ≥n when the read hit its limit, so a busy day reports a floor rather than describing a prefix as if it were the whole day.

Spend (month to date)

Against your monthly cap, with a bar that turns amber at 80% and red past 100%. The figure is the same one budget caps enforce, so the two can never disagree. With no cap set it falls back to the 24-hour total and says so.

Success rate (24h)

Of runs that reached a verdict. Cancelled runs leave the denominator entirely — a person pressing Stop is not a failure. With nothing decided it reads “—” rather than congratulating you on 100%.

Avg latency (24h)

Mean wall-clock time per model call, over the runs that recorded one.

Activity and model mix#

Activity charts hourly run volume for the last 24 hours across every agent and swarm. Model mix beside it ranks models by tokens, so you can see where the spend went before opening anything.

What you’re running#

A grid counted from this deployment’s own tables, covering both halves of the platform: agents, swarms and knowledge bases on one side; ETL pipelines, the lakehouse, SQL models, ML models, dashboards, workflows, data monitors, metrics and integrations on the other. A capability with something in it shows the count and, when relevant, what is wrong with it — “5 watching · 1 open”. A capability with nothing in it says so and offers the way in, rather than showing a zero.

Spend & usage — by person, team or organisation#

The Spend & usage panel attributes model cost, so it can be charged back rather than only totalled. Two pickers control it.

The picker offers only the scopes you may actually use. If you are in no team, “My teams” is absent rather than present-and-broken — and a scope you are not entitled to is refused, never quietly answered with your own numbers under someone else’s label.

The time range covers the last 24 hours through year to date. Windows are half-open and in UTC, so a run landing exactly on a boundary is counted once and the same dashboard reads the same from any timezone.

Cost comes from the same column the budget caps read, so a figure here and a budget alert can never disagree about what someone spent.

Recent runs#

The last six executions across your workspace, each with the agent name, model, latency, cost, and a success/error indicator. View all opens the full run history at /traces, where every run can be expanded into its complete trace — see Logs & traces for how to read one.

What to do on a new workspace#

A deployment with nothing built and nothing run gets a different page: an ordered checklist instead of a console full of zeroes, with each step ticked from real state rather than from a flag. Only the next unfinished step carries a button. The order below is the same one, and it is dependency order rather than feature order.

1. 1

   Connect a model provider

   Integrations. Until you do, calls run on the operator's shared fallback key — fine for a first look, wrong for anything real. See Models & providers.
2. 2

   Add data or documents

   Data Catalog for rows, Knowledge Base for prose. An agent with neither is just a chatbot.
3. 3

   Build one agent

   Agent Builder — name, prompt, model, one or two tools.
4. 4

   Run it and read the trace

   Logs & traces. This is the habit worth forming early.
5. 5

   Before anyone else joins

   Turn off public signup and set budget caps — Access control and Budgets.