Dashboard

TL;DR: /admin is the overview — one screen for the whole instance over the last 24 hours, 7 days or 30 days. Ten headline tiles compare the window with the one before it, five charts show it over time, two leaderboards rank projects and groups by spend, one table ranks models, and a queue shows the approvals and anomalies waiting on a person. The page refreshes itself every half minute while it is open. Every figure links one click deeper.

Who sees it

The overview is for the console roles, admin and platform_admin. Anyone else who opens /admin is sent to their profile, or — if they participate in a marketplace — to the scoped pages that participation opens. See Create & Manage Users → Roles.

The window

The 24h / 7d / 30d buttons in the page header choose the window (24 hours is the default; the choice is in the URL as ?preset=, so it can be bookmarked). Changing it redraws the page without a reload.

Each tile's note compares the window with the one immediately before it ("vs previous 7d"). When the previous window carried too few requests to compare against, a note says so and no tile shows a percentage.

Live refresh

While the tab is visible the page fetches its own figures every 30 seconds and updates them in place — tiles count to their new value, charts keep your hidden series and focus, changed table cells flash. The Live indicator in the header shows how old the figures are; Pause stops the refresh, and it stops by itself when your session ends or the fetch keeps failing. Motion is switched off under your system's reduced-motion setting, and the page works without JavaScript — it just does not refresh.

The live figures come from GET /admin/api/overview, which renders the same markup as the page from the same queries, so a refreshed number can never disagree with the one on first load.

Headline figures

Tile	What it counts	Opens
Requests	Gateway requests in the window	Request log
Spend	Gateway cost in the window	Analytics → Cost
Active people	Distinct accounts that made a request	Users
Conversations	Distinct gateway conversations	Analysis → Conversations
Tokens	Input and output tokens; the note gives the share read from cache	Analytics → Models
Tool calls	Tool calls across every client	Analytics → Tools
p50 latency	Half of routed requests finished faster than this	Traces
p95 latency	95% of routed requests finished faster than this	Traces
Error rate	Failed requests over all requests	Request log, failed only
Denied	Requests recorded as rejected — refused before routing	Request log, rejected only

While governance, safety and quotas run in warn mode, Denied counts only what is still refused — for example a model the gateway does not route — not what warn mode would have blocked. Those are recorded as warnings; see Governance Modes.

Over time

Five charts, bucketed to the window:

• Requests, conversations and people
• Tokens and cache reads
• Tool calls, errors and denials — with rejected requests as their own series
• Latency per bucket — p50 and p95, captioned with the worst bucket of each
• Spend by model and Requests by client — the top series, the rest as Other; requests for a model the gateway does not route appear as Unrouted

Where the spend goes

Top projects by spend and Top groups by spend sit side by side and share one layout: requests, spend and share of the window's spend. Attribution is exclusive — each person counts once, under their primary project or group — so each board adds up to the whole. People with no primary project or group appear as Unattributed. All opens the full list; a row opens that project's or group's page with its own usage tab.

Most used models

One row per model the gateway served in the window, ordered by requests, with share of traffic, spend and median latency. A model the gateway refused is marked REJECTED. All models opens the Analytics Models tab, which reads the same rows.

Waiting on a person

• Pending approvals — tool calls held by a require_approval policy. Opens the approvals queue at /admin/governance/approvals.
• Open anomalies — usage the anomaly detector flagged in the last 24 hours, listed below the tile with metric, window, observed value, baseline and severity. Anomalies are shown here and written to the server log; no notification is sent anywhere else. See Audit & Observability.

Getting around

The header buttons open the Request log and Analytics. The sidebar groups the console:

Group	Pages
Activity	Overview, Analytics, By person
People	Users, Groups, Projects, Roles & permissions, Access control, Devices & enrolments
Analysis	Conversations, Skills, Reports, Versions
Governance	Governance, Approvals, Secrets, Quotas
Platform	Configuration, Lifecycle, Code sync, Marketplaces, Plugins, Skills, MCP servers, Gateway
Account	Profile, Connect, Connectors, My conversations
Developer	Requests, Sessions, Conversations, Traces, Tools & artifacts

The same figures are available from the CLI for scripting and export:

systemprompt analytics overview --since 7d
systemprompt analytics overview --export overview.csv
systemprompt analytics costs breakdown --by model --since 30d

Check it yourself

1. Open /admin, choose 7d, and note Requests.
2. Click the tile. The request log opens on the same window; its count matches.
3. Run systemprompt analytics overview --since 7d and compare the request total. A small difference is requests that landed between the two reads.