Monitoring

AgentData gives operators two windows into a running system: the Dashboard (usage, cost and a health snapshot) and Monitoring (a live activity feed, split by subsystem — the Monitoring tab of the app's Settings & Monitoring section). Admins see their own tenant; super-admins see every tenant with a tenant filter.

The Dashboard

The Dashboard is the at-a-glance view. A time-range picker (24h / 7d / 30d / all) and — for super-admins — an all-tenants view with a per-tenant filter rescope everything on it.

It shows:

• the discovery pipeline and model snapshot — connected sources, pending vs confirmed entities,
• the LLM balance banner (for metered tenants) with top-up request and coupon redemption,
• usage & cost broken down by query type, by model, and by user — with pie charts — for the selected window,
• query statistics — totals, distinct users, volume by source (UI vs MCP) and over time,
• recent sources and scans, and (super-admin) the system log with CSV/text export.

The Monitoring hub

Monitoring is a live feed with a tab per subsystem. Each tab has a super-admin tenant filter and a live-refresh toggle.

Flows

Execution logs for every flow run — timestamp, level (info / warn / error), step description, message and detail. Filter by level or free-text search. This is where you confirm a scheduled load succeeded, or find why one failed.

Queries

Every query executed across the tenant — from the UI, from MCP clients, and from the Data API. Columns show when, source, user, the question, status and row count, with filters for new / reused / error and by source (UI / MCP / Data API). It's the operational counterpart to the History tab.

Feedback

The open feedback queue — user remarks on saved queries awaiting a BA's review — surfaced here as well as inside the Query view, so operators can see what needs attention.

Data API

The health of your published Data API services: queue depth (pending / running / done / failed / dead), a per-service table with call counts, error counts and average / p95 latency, and the most recent calls. Use it to spot a slow or failing endpoint before consumers complain.

Streaming

The state of your event streams: the broker mode (internal or cloud) and, per stream, whether it's enabled, its target table, and when it last received an event. (Per-event logs aren't retained — this is the configuration-and-liveness view.)

MCP traffic

MCP access has its own monitor under Settings → MCP: calls broken down by client app (claude.ai, ChatGPT, Cursor, CLI…) and by user, the active API keys and their scopes, and any failed-auth or lockout events. See MCP server and Security.

Reliability & self-healing

Beyond the dashboards, an optional watchdog runs alongside the stack: it health-checks each component on a schedule, attempts to self-heal a stalled service (restart), and raises alerts (email / Slack) when something needs attention. It's config-driven and runs under launchd (macOS) or systemd (Linux) — see Configuration.

Who can see what

Role	Dashboard	Monitoring	Scope
viewer / editor	own usage	—	their own activity
admin	own tenant	✓ own tenant	one tenant
super-admin	all tenants + filter	✓ all tenants + filter	every tenant

Next steps

• Data movement — the flows, streams and endpoints you're watching
• Administration — budgets, retention and configuration behind these numbers