Observability

Observability

The group opens on LLM logs and is visible with Logs.View.

Home (/workspace/inicio)

First screen after login. An executive read of the month, not an analysis panel. Everything comes from real APIs; with no data, it shows empty states. Subtitle on the screen: what went through your organization's AI governance in the last 30 days.

• Four KPIs over a fixed 30 day window: month spend, governed conversations, active agents, generated savings.
• Charts for activity per day and spend per provider.
• Cards: who uses AI the most (top 5 users by cost), recent conversations, workspace inventory (connected providers, catalog models, virtual keys, observed agents).
• Agent and savings KPIs come from the management endpoints /api/atlasberg/agentes and /economometro; a failure renders as module unavailable in this deployment, never as a page error.
• Nothing to configure. Read-only screen with a shortcut to the full dashboard.

Dashboard (/workspace/dashboard)

The full analysis panel, with tabs for overview, model rankings, provider usage, dimension rankings and MCP.

• All filter state lives in the URL: period (default 30d), start and end, timezone, virtual keys, providers, models, status, routing rules, stop reasons, cache hits, metadata filters, users, teams, customers, business units, apps, and chart type per metric (bar or line).
• Per-tab export via the export popover.

LLM Logs (/workspace/logs)

Table of LLM requests with a volume chart, a filter sidebar, a per-log detail sheet and a per-session detail view.

• Filters by provider, model, parent_request_id and other facets, kept in the URL.
• Deleting logs requires the Logs.Delete permission; revealing protected content requires Logs.Reveal.

MCP Logs (/workspace/mcp-logs)

MCP tool execution logs.

• URL filters: tool names, server labels, status, virtual key ids, content search.
• Deleting requires MCPLogs.Delete.

Connectors (/workspace/observability)

Configure observability connectors to external platforms.

• Available in every build: OpenTelemetry, Prometheus, Maxim.
• Enterprise builds only: Datadog, BigQuery, Kafka, Pub/Sub. A non-enterprise build does not even offer those cards.
• Each connector has its own configuration form.

Logs Settings (/workspace/config/logging)

Gateway logging configuration. The view reads the client config and the enforcement block, the content lock measured over all destinations: logging, each OTel profile, semantic cache.

• Visible field: logging_headers, the list of logged headers, edited as comma-separated text.
• Changes may require a restart; the screen indicates it.

Costs

In the console menu the seven screens below form the Costs group (URL /workspace/custos, always visible), described as spend, chargeback, budgets and efficiency in one place. Menu order: Executive panel, FinOps, Overview, Efficiency, Soft budgets, Savings meter, Copilot. None of them declares an RBAC resource.

Executive panel (/workspace/painel-executivo)

Spend drill-down from company to organizational unit to team to user, with a breadcrumb trail. Numbers at each level add up to the level above. Data comes from GET /api/atlasberg/executivo.

• Window: presets of 30 days, 90 days and everything, plus two date fields. With no filter the label reads whole history.
• Signed PDF export: the export button downloads atlasberg-executivo.pdf; a .txt export also exists. The envelope is signed with the deployment key and verifiable offline with atlasberg-verify.
• Ceiling per team: the percent-of-ceiling column comes from governance (monthly team budget). Only teams have a ceiling; ceilings are not summed upward because a group ceiling does not exist in governance.

Sub-screen: Organizational units

At /workspace/painel-executivo/unidades. CRUD of units: create a unit with a name and an optional parent, rename, and map or remove teams, via the /api/atlasberg/executivo/unidades routes. Moving a team only fires on the confirmation button, not on the select.

FinOps: chargeback (/workspace/chargeback)

Cost allocation by owner (virtual key), provider and model, aggregated from the audit trail, the same source as the compliance report. KPIs: requests, tokens, total cost, and not served (admitted but no response).

• Owner resolution: the trail delivers a pseudonym (a digest of the virtual key); the console matches the digest against the governance virtual key inside the user session, which is allowed to see both ends. Without a secure context (https or localhost) the browser does not expose crypto.subtle and the owner stays as the digest.
• Configuration: a period filter with two date fields and a refresh button.

Overview (/workspace/custos)

One screen, one question: how much are we spending, on what, and is it under the ceiling. Spend and ceiling side by side.

• KPIs: period cost, requests, cost per thousand requests, teams above 90 percent of their ceiling.
• Charts: daily cost per model (five named series plus others) and month budget with a consumption gauge per team ceiling.
• Each card declares its source in the footer: the observability database (logs.db) or governance (config.db).
• Configurable: period selector of 7, 30 or 90 days. The budgets themselves are configured on teams in governance, not here.

Efficiency (/workspace/custos/eficiencia)

Cost per thousand requests, per team. Click a bar to see that team's model mix, which is where the difference explains itself.

• Source: one scope per team via GET /api/atlasberg/executivo.
• Configurable: period of 7, 30 or 90 days.

Soft budgets (/workspace/orcamentos)

Budgets with a soft cut. An opt-in monthly limit per organizational unit, team or user. When it blows, requests in that scope switch to the model you choose. No request is ever denied: the opposite of Budgets & Limits under Models, which blocks. Managed via the /api/atlasberg/orcamentos routes.

• Rules table columns: scope, limit, month spend with progress bar, target model, degradations, blocked, active toggle, delete action.
• New rule form: scope type (organizational unit, team or user; default team), scope (a select fed by the executive panel directory, degrading to a text input if the directory fails; for a user, the value of the user field sent in the request body), monthly limit in USD (minimum 0.01), target model (a select from the model catalog, avoiding a silent typo that would only fail on the first degradation), optional target provider (default keeps the request provider), and an enable-now checkbox, checked by default.
• Behavior documented on the screen: each degradation is recorded in the trail as budget.degraded; turning a rule off returns traffic to the original model immediately.
• If the target model is not allowed for the virtual key, the cut prefers not to degrade and warns: a rule blew and could not degrade, counted in the blocked column.

Savings meter (/workspace/economometro)

A passive cost waste meter: the gateway observes tokens, declared ceilings and context repetition in the traffic it already governs, and points to where money is leaking. It never blocks and never alters any request. Data from GET /api/atlasberg/economometro, live every 10 seconds.

• Shows: estimated savings in the window (anchored to real prices from the model catalog; without a catalog price the estimate stays at zero and the screen says so), active findings, requests analyzed, finding cards and a daily traffic series.
• The four detectors: giant prompt, exaggerated max_tokens, repeated context, expensive model on a trivial task.
• Off state: when disabled by policy, the screen says the meter is off and the data below stopped accumulating.
• Nothing to configure on this screen; the on-off switch comes from policy.

Copilot (/workspace/copilot)

License cost imported from Microsoft administrative APIs. It does not go through the gateway: this is money and ownership, not content. The copilot license module gates the import, not the reading: the server always answers the whole panel and adds the fact modulo_licenciado plus, when the module is missing, a modulo_aviso notice. The screen shows that notice above the KPIs and keeps rendering the panel with the inventory already imported; without the module there is no new import and no new connection.

• KPIs for period cost, seats, idle seats (no activity in 30 days) and idle cost (paid license with no use).
• Cards per surface: GitHub Copilot, Microsoft 365 Copilot, Copilot Studio, billed per seat or per consumed pack (included, consumed, overage, remaining).
• Per-team chargeback table.
• Before the first import the surface and per-team lists come back empty (never null), so the screen says nothing imported instead of failing.
• Data freshness: a data-as-of stamp next to every number, with expected, late and stale states. The APIs close per day, so the inventory never describes today.
• Nothing to configure here; configuration lives in Governance, Copilot access.

About Atlasberg Platform

Atlasberg Platform is the control layer between a company and every AI model: each request is authenticated with a virtual key, filtered by policy and DLP, routed to the right provider and written to a hash-chained audit trail. It exposes an OpenAI-compatible API, so applications only swap the base URL.

The same artifact runs in Atlasberg Cloud, in your VPC, on-premises or fully air-gapped, and is priced by capacity and modules, never per seat.

This page is part of the official documentation. To talk to the engineering team, write to [email protected] or use https://atlasberg.com/contato. Answers come within one business day.