The two-star icon in the top right of Grafana Cloud is the Assistant, and the short version is that it is very good. It will find out what data you are actually receiving, build a dashboard for it, and fix the dashboard when you tell it what is wrong - in plain language, without you learning a query syntax or a panel type.
For anyone who is competent at monitoring but does not live in Grafana, that removes a genuine barrier. It is also closed, metered, and on the self-hosted path it sends your queries to Grafana Cloud whether or not the rest of your install does. Both halves are worth understanding before you build a habit around it.
Three things, in increasing order of how much they change your working day.
It answers questions about Grafana, which is the least interesting capability - smart documentation with a chat box. Useful, unremarkable.
It inspects your instance. Ask what metrics are arriving and it goes and looks. This is where it stops being a chatbot: a question like "there should be data coming in from my local Alloy collector" gets checked against reality, and it will come back with the number of series and what they contain. For the specific and common situation of data arriving somewhere you cannot find in the UI, this is faster than navigating.
It builds things. "Create a dashboard for that data" produces a real dashboard with real queries, organised into rows, and it will iterate - noticing a panel it rendered badly and reworking it. It takes correction well: a legend showing only probe names when several lines shared a probe was fixed by describing the problem in a sentence, with the Assistant working out that the metric carried a URL label and rewriting the legend accordingly.
It also has specialised modes - dashboarding, alerting, k6 scripting, knowledge graph - so you can point it at the right competence rather than hoping.
Underneath is an MCP server built by Grafana that exposes your instance, with an LLM on top translating plain English into calls against it. That architecture explains both the strengths and every limitation that follows.
The strength: because it holds real tools against your live instance, it can inspect and act rather than just describe. An MCP server alone would give you the actions without the plain-English interface; an LLM alone would give you confident descriptions of a Grafana it cannot see. The combination is what makes it useful.
The consequence: LLM inference is not free, so the Assistant is not free, and it is not part of the open-source Grafana project.
| Free tier | Limit |
|---|---|
| Active AI users | 3 per month |
| Tokens | 40M per user per month |
| System-initiated usage | 25M tokens per month |
| Support | Community only |
From the Grafana Cloud pricing page, checked 2026-08-28. Pro and Enterprise raise these substantially.
Forty million tokens a month per user is generous for interactive use - you would have to work at exhausting it by asking questions and building dashboards. The binding limit for most teams is three active AI users, which is fine for an individual or a small team and stops being fine the moment you want it available to everyone who touches dashboards.
Self-hosted Grafana can use the Assistant. Go to Administration → Plugins and data → Plugins, search for assistant, install the Grafana Assistant plugin, and connect it to a Grafana Cloud account. You still need that account even though your Grafana is not in the cloud.
Everything you type goes to Grafana Cloud. The plugin is a client. Your prompts, and the instance context needed to answer them, are processed in the cloud and the result comes back. Running Grafana on your own hardware does not keep Assistant traffic on your own hardware, and the UI does not labour the point.
Whether that matters depends entirely on what your Grafana can see. Asking it to tidy a panel is uninteresting. Asking it to investigate an incident means the questions - and the context it gathers to answer them - describe your infrastructure, your naming, and quite possibly what is currently broken and how. For most teams that is an acceptable and unremarkable trade. For anyone under data-residency constraints, or monitoring systems where the metric names alone are sensitive, it is a decision that should be made deliberately rather than by installing a plugin.
There is currently no self-hosted inference option. Cloud is the only path.
You cannot bring your own model. There is no way to point it at your own API key and have it drive Grafana's MCP server with a model you already pay for. That is a reasonable product decision - the quality comes partly from tuning the model and the tools together - but it means the capability is only available on their terms and at their price.
You cannot compose it with anything. The MCP server is reachable only through the Assistant inside Grafana. You cannot attach it to your own client alongside other MCP servers, which is exactly what you would want for the questions that span systems: correlate this Grafana alert with that deployment and the ticket that opened. Each vendor's assistant sees only its own product, and the interesting questions never respect those boundaries.
That is the structural cost of a vendor-hosted assistant, and it is worth weighing against the alternative shape - an open MCP server you host and point any client at, which is what the Zabbix ecosystem has. More setup, less polish, and it composes.
Use it. If you are not fluent in Grafana it collapses a learning curve into a conversation, and the dashboards it produces are real ones you can read and learn from. The free tier is enough to find out whether it fits how you work.
Just hold two things in mind: the three-user limit is the one that will bite as soon as it becomes popular on your team, and on self-hosted installs it is a cloud service wearing a local badge. Neither is a reason to avoid it. Both are reasons to decide rather than drift.
Eleven minutes, including the dashboard being built from a single sentence and the walk through the self-hosted plugin path.
The self-hosted-versus-cloud line is rarely about features. It is about which data is allowed to leave, who is asking, and what you have to be able to prove. If you are working that out for a monitoring stack, or trying to keep a hybrid arrangement coherent, that is the kind of thing I help with.