Dashboards

Build observability dashboards by describing what you want in plain language. Rounds discovers your metrics, picks visualizations, validates the queries, and ships a working dashboard grounded in real values.

What you can do

• Generate from a prompt — "create a dashboard for HTTP latency" → ~5 panels with grounded p50/p95/p99, request rate, error rate.
• Iterate by chat — "add a panel showing 5xx errors by handler", "split the latency panel by method", "remove the request-rate panel".
• Manage layout — rearrange, resize, retitle panels via chat or drag-and-drop.
• Add variables — "add a service dropdown" creates a template variable backed by a label query.
• List & open — dashboard.list returns dashboards filtered by folder/tag; click-through to open in the workspace.

How to use it

Create a new dashboard

In the chat panel:

Create a dashboard for HTTP latency

Rounds runs the orchestrator agent through a multi-step plan:

1. datasources.list — find available metrics backends
2. metrics.metric_names + web.search — discover relevant metric names + best practices
3. metrics.metadata / metrics.labels — understand the schema
4. metrics.query (parallel) — sample real values for grounding
5. metrics.validate — confirm each query parses and returns data
6. dashboard.create + dashboard.add_panels — build it
7. navigate — open the new dashboard for you

You'll see the streaming step trace as it runs.

Edit an existing dashboard

Open a dashboard, then in the chat:

Add a panel showing the 5xx error rate per handler

Or use direct UI controls — click the panel menu to edit/duplicate/remove. Both paths use the same underlying tools (dashboard.add_panels, dashboard.modify_panel, dashboard.remove_panels).

Add a template variable

Add a service variable populated from the service label

Generates a metrics.label_values query and wires the variable into all panel queries that use that label.

Examples

Prompt	Result
Create a dashboard for Redis health	Connections, ops/sec, memory, evictions, hit rate
Add p50/p95/p99 latency panels grouped by route	3 stat panels + 1 timeseries panel
Split the requests panel by status code class	Modifies existing panel to add status_code_class legend
Remove all panels with zero data	Iterates panels, queries each, deletes empty ones

Limits

• Generation is grounded in your current metrics. Cardinality bombs (label combos with millions of series) are filtered out automatically — limit 20 series per query result.
• The model picks panel types heuristically. You can override with prompts like "use a heatmap for that latency panel".
• Variable queries with high cardinality (>1000 values) get truncated; use regex filters to narrow.
• Dashboards are JSON-compatible with the Grafana schema; you can export / import between systems.

Related

• Datasources — what metric backends are supported
• Chat & agents — how the dashboard agent is wired
• Permissions — folder-scoped grants for who can view/edit