Ask a question, get the query it used.
Every answer names the query it ran, so you can check it rather than trust it.
The Dashboards tab puts your Grafana dashboards beside a chat on a persona built for them. It can enumerate dashboards and report each title and uid, inspect datasources before querying so it targets the right one, query Prometheus for metrics and Loki for logs, and cite the dashboard, datasource and query behind every answer. It can also edit dashboards: fetch the current JSON by uid, change it minimally, summarise the concrete before and after, and only push the update once you confirm.
Citing the query is the point
The persona is required to name the dashboard, datasource and query behind each answer. That converts a plausible sentence about latency into something you can paste into Grafana and check yourself, which is the only thing that makes an agent trustworthy against production data. An uncited number from a language model about your error rate is worth nothing.
Edits are confirmed before they land
Changing a dashboard means fetching its current model by uid, editing that JSON as little as possible, then summarising the change in concrete terms, which panels, which queries, before and after, and waiting for you. The uid stays stable across an edit, and the persona is instructed never to delete or overwrite a dashboard you did not ask about.
Viewing without leaving the window
Dashboards render in a live kiosk-mode frame, or as a server-rendered image refreshed on an interval when frames are blocked by your Grafana's security headers. The image path needs Grafana's image-renderer plugin installed on the server, which is the usual reason it is unavailable.
It refuses to guess
With no Grafana connected there are no mcp__grafana__ tools, and the persona is instructed to say so and point at the connect card rather than inventing metrics. Fabricated observability data is worse than no answer, because it looks exactly like a real one.
Questions
Related
- From a Production Symptom to a Code ChangeNot an automated root cause engine. One window that holds both halves of the investigation.
- Connect Grafana to a Coding AgentThe tools are Grafana's own MCP server, not a reimplementation of the Grafana API.
Last updated