AI Capabilities in the Grafana Stack
For years, SRE teams have relied on dashboards, alerts, and manually written queries to understand what was happening inside their systems. But modern infrastructure has become distributed, complex and noisy for SREs to constantly monitor graphs and correlate events manually.
This is where AI-assisted observability starts making sense. The shift happening right now is from dashboard-based observability to intent-based observability.
Instead of asking questions like:
- “Which dashboard should I open?”
- “Which PromQL query should I write?”
- “Which service is failing?”
Engineers can now simply ask:
- “Why did latency spike in production?”
- “Which deployment caused this issue?”
- “What changed before error rates increased?”
This is also closely tied to the rise of AI-powered SRE workflows. AI is now being used to summarize incidents, correlate telemetry automatically, detect anomalies, assist with root cause analysis and help engineers navigate large observability datasets faster.
Observability platforms are evolving quickly to support this shift. One of the biggest examples of this is Grafana's recent push into AI capabilities across both Grafana Cloud and Grafana OSS. Take this with a grain of salt. Go through the next sections to discover more about the limitations.
In this blog, we'll look at the new AI features introduced in Grafana, where they fit inside the PLG (Prometheus, Loki, Grafana) stack, how setup differs from Cloud to OSS, and whether these features actually improve day-to-day observability workflows.
Understanding the PLG Stack
Before we talk about AI features, it is important to understand where these capabilities are added in the observability stack. The PLG stack refers to Prometheus for metrics, Loki for logs and Grafana for visualization and analysis. When you add Tempo and Mimir, it is popularly known as LGTM (Loki, Grafana, Tempo, Mimir).
Each layer has a different responsibility.
| Layer | Purpose |
|---|---|
| Prometheus | Metrics collection and alerting |
| Loki | Centralized logs |
| Tempo | Distributed tracing |
| Grafana | Dashboards, correlations, querying, investigations |
The important thing is that AI is not replacing these systems. Instead, AI is being added on top of the observability stack to make telemetry easier to consume and investigate. Grafana is essentially becoming the AI interaction layer for observability data.
New AI features in Grafana 13
Recently, Grafana introduced several AI-powered capabilities aimed at improving investigation workflows and reducing the operational overhead of observability. Some of these features are available directly in Grafana Cloud, while others can also be configured in self-hosted OSS deployments.
| Feature | Grafana OSS | Grafana Cloud |
|---|---|---|
| AI Assistant | Available through connection with Grafana Cloud (Paid feature) | Available natively |
| Customizing Dashboard Templates with Grafana Assistant | Available through connection with Grafana Cloud (Paid feature) | Available natively |
| AI-Assisted Root Cause Analysis | Not available, requires data ingested directly into Grafana Cloud. | Available natively |
| Anomaly Detection and Forecasting | Not available, requires cloud-hosted telemetry storage | Available natively |
| Fast Onboarding with Prompts | Available through connection with Grafana Cloud | Available natively |
| Skills and Repeatable Workflows | Available through connection with Grafana Cloud | Available natively |
| Grafana MCP Server | Available natively, can connect external AI agents that are MCP-compatible clients | Available natively |
AI Assistant
The AI Assistant is one of the most visible additions. It allows users to interact with observability data using natural language instead of manually navigating dashboards and writing queries.
Example use cases:
- Ask questions about spikes in latency
- Query logs without writing LogQL manually
- Understand what a dashboard panel represents
- Generate PromQL queries from prompts
- Summarize incidents or telemetry patterns
This changes the onboarding experience significantly for newer engineers who may not be familiar with PromQL, Loki queries, or Grafana dashboards. Instead of learning query syntax first, they can start by asking questions in plain English.
Customizing dashboard templates with Grafana Assistant
Traditionally, dashboard templates help teams get started quickly, but they often require manual adjustments because real production environments rarely use the exact same metric names, labels, datasource structures or naming conventions.
Grafana Assistant attempts to reduce that friction by helping adapt templates using natural language interactions. This makes dashboard onboarding significantly easier, especially for teams working with large observability environments, inconsistent telemetry naming or multiple datasource integrations.
AI-Assisted root cause analysis
Grafana has also introduced AI-assisted investigations for root cause analysis workflows.
This feature attempts to correlate metrics, traces, logs, deployments and infrastructure events to help narrow down possible causes during incidents. Rather than jumping across multiple dashboards manually, investigations become more guided.
For example:
- a latency spike can be connected with a deployment event,
- correlated with error logs,
- and traced back to a failing dependency.
This is one of the strongest practical AI use cases in observability because incident response is usually where engineers spend the most time.
Anomaly Detection and Forecasting
Another area where AI fits naturally is anomaly detection. Traditional alerting depends heavily on manually defined thresholds, but static thresholds often fail in dynamic systems. Grafana's AI-assisted anomaly detection tries to identify unusual behavior patterns automatically instead of relying entirely on fixed alert rules.
Forecasting features also help predict storage growth, metric trends, capacity exhaustion and traffic patterns. This becomes useful for proactive operations and infrastructure planning.
Fast onboarding with Prompts
One underrated benefit of AI in observability is onboarding speed. Large observability environments can become difficult to navigate hundreds of dashboards, thousands of metrics, multiple data sources and inconsistent naming conventions.
AI-assisted prompts reduce the friction significantly. Instead of searching manually, engineers can ask questions to understand the setup. This lowers the learning curve for new team members.
Skills and Repeatable Workflows
Grafana has also started introducing concepts similar to reusable AI workflows or “skills.” These are somewhat comparable to reusable prompt workflows in tools like Anthropic's Claude. The idea is to standardize common investigation tasks.
Examples:
- deployment analysis workflow,
- database incident investigation,
- Kubernetes troubleshooting,
- API latency analysis.
Instead of manually repeating the same steps during every incident, workflows can become semi-automated and reusable. This is especially useful for SRE teams handling recurring operational patterns.
Grafana MCP Server
Another interesting addition to Grafana's AI ecosystem is the Grafana MCP Server supported from version 9.0. MCP (Model Context Protocol) allows external AI agents to interact directly with Grafana and observability systems in a more structured way.
Instead of relying only on Grafana Assistant inside the Grafana UI, MCP enables tools like Claude, Cursor and other MCP-compatible AI clients to query dashboards, inspect observability data, generate visualizations, troubleshoot incidents and interact with Grafana conversationally.
This opens up a more flexible AI workflow where Grafana becomes part of a broader AI-agent ecosystem rather than a standalone observability UI. For self-hosted environments, this is especially interesting because it provides a more open integration model compared to some of the more cloud-coupled Assistant workflows.
Demo: How to use AI features in Grafana Open Source
To test the newer Grafana Assistant capabilities, I used a hybrid setup where a self-hosted Grafana OSS instance was connected to a free Grafana Cloud account.
This setup is important because many of the newer Assistant workflows in Grafana rely on Grafana Cloud as the backend AI service, even when using a self-hosted OSS deployment.
1. Running Grafana OSS locally
The first step was running Grafana OSS locally using Docker.
docker run -d --name grafana -p 3000:3000 grafana/grafana:13.0.1
After starting the container, Grafana was accessible locally at: http://localhost:3000
I used Grafana OSS v13 because the newer Assistant ecosystem and AI integrations are designed around the Grafana 13 release cycle.
2. Creating a Grafana Cloud account
Next, I created a free Grafana Cloud account and created a Cloud stack.
The free tier was enough for experimenting with the Assistant integrations and OSS connectivity. To get more details on pricing and limitations you can refer to Grafana Pricing.
3. Installing the AI Plugins in Grafana OSS
To enable the AI integrations locally, I installed the Grafana Assistant plugin.
The plugin was installed directly inside the Grafana container with command
docker exec -it grafana grafana cli plugins install grafana-assistant-app
After installing the plugins, Grafana was restarted.
4. Connecting Grafana OSS to Grafana Cloud
Once the plugins were available, the next step was connecting the OSS deployment to Grafana Cloud.
5. Exploring the Assistant Interface
The setup showed how AI-assisted workflows can simplify dashboard navigation and telemetry exploration using prompts. At the same time, many of the newer AI capabilities still depend heavily on Grafana Cloud services rather than being fully standalone OSS features.
Hands-On: Trying out the AI features
To test the AI capabilities in a practical way, I decided to use the grafana-testdata-datasource.
Using AI Assistant to build and customize dashboards
The first thing I tested was using Grafana Assistant for dashboard creation and customization. Instead of manually creating panels and writing queries, the workflow became more prompt-driven.
For example, I asked the Assistant to create a dashboard and add panels using the Grafana test data source. The Assistant also helped customize templates and adapt dashboards more interactively.
Fast Onboarding Through Prompt-Driven Workflows
I also tested how AI-assisted workflows could help with onboarding. Instead of manually exploring dashboards and learning query syntax, new engineers could ask questions and understand the setup. This made the observability setup feel easier to navigate, especially for engineers unfamiliar with Grafana dashboards or query languages.
AI-Assisted troubleshooting and investigations
After creating the dashboard and testing onboarding workflows, I moved to simple troubleshooting scenarios using the Grafana test data source.
The demo dashboard included latency spikes, fluctuating error rates and changing infrastructure usage patterns. This allowed testing how the Assistant could help during investigations using prompts such as:
- “Why is API latency increasing?”
- “Correlate error rates with CPU usage.”
- “What changed before the latency spike?”
The main value was not automatic incident resolution, but reducing the effort involved in investigations. Instead of manually switching between dashboards, writing queries and correlating telemetry, the workflow became more prompt-driven and easier to navigate.
Grafana MCP Server
Alongside Grafana Assistant, I also experimented with the newer Grafana MCP (Model Context Protocol) server approach with my Claude Desktop Agent. Instead of relying on Grafana Cloud-hosted AI features, MCP allows external AI agents to interact directly with a Grafana OSS instance.
Grafana already provides detailed setup documentation for the MCP server and supported clients in its official documents page. The official documentation can be followed here:
Once connected, Claude was able to directly inspect dashboards, datasources and observability resources from the Grafana OSS environment using natural language prompts.
Example Prompts Tested
The prompts I tested included:
- “List my Grafana dashboards”
- “What datasources exist in this Grafana instance?”
- “Create a dashboard for CPU and memory monitoring”
The interesting part was how different the workflow felt compared to traditional dashboard navigation. Instead of opening Grafana manually, browsing folders, writing PromQL queries, selecting visualizations and correlating metrics manually, the interaction became significantly more intent-driven. That distinction sounds small initially, but it changes how observability systems are explored, especially for onboarding engineers or quick investigations.
Key Observations
The most useful part of Grafana's AI features was not automation, but reducing investigation overhead.
The Assistant workflows made dashboards and telemetry easier to navigate, especially for onboarding and quick troubleshooting scenarios. The MCP server approach was particularly interesting because it allowed external AI agents like Claude to interact directly with Grafana OSS. Unlike the current Grafana Assistant model, this setup felt more modular and OSS-friendly since the AI layer could be replaced independently.
At the same time, many of the polished Assistant workflows still depend heavily on Grafana Cloud services. While OSS integrations exist, the overall Assistant ecosystem is not yet fully self-hosted or model-agnostic.
Conclusion
Grafana's AI capabilities show where observability tooling is heading: toward conversational and investigation-driven workflows instead of purely dashboard-driven operations.
Today, the biggest value is not fully autonomous operations, but faster onboarding, simpler investigations and easier navigation across large telemetry environments.
The Grafana Assistant experience currently remains tightly coupled with Grafana Cloud, but the MCP ecosystem introduces a more flexible direction where external AI agents can interact directly with self-hosted Grafana environments.
Originally published at https://www.cloudraft.io on May 26, 2026.
Top comments (0)