1
Sign up and confirm your email
Open the sign-up page and either use a social sign-in button or fill in Name, Organization, Email and Password (at least 12 characters), then select Create account.Signing up creates a new organization with you as its administrator. If your email domain already belongs to an organization, sign-up sends that organization’s administrators a request to admit you instead.Open the confirmation email and follow the link. Investigations stay disabled until your address is confirmed.
2
Connect an alert source
In the sidebar, select Integrations. The Webhooks tab is open by default and shows a card for each supported source. Under Grafana, copy the Endpoint URL. It has this shape, with your organization’s own token at the end:Treat the URL as a credential: anyone who has it can post alerts into your organization.Then, in Grafana:
- Go to Alerting, then Contact points, and select Add contact point.
- Choose Webhook as the integration type.
- Paste the endpoint URL into the URL field and set the HTTP method to POST.
- Save, then assign the contact point to a notification policy so your alert rules use it.
severity label sets the severity in SRE Agent: critical, high, warning and info map directly, and an alert with no severity label becomes medium.3
Connect an AI provider (optional)
SRE Agent provides shared AI by default, so investigations work without this step. To use your own
provider, go to Settings, then AI Providers, then Add Provider. After you save it,
press Test on the provider card to check the key. Connect your
data covers it, including how to keep your data on your
own providers only.
4
Add a data source
The agent can only investigate with what it can read. Still in Settings, open the Data Sources tab and select Add Data Source. Enter a Name, choose the Type (for example Prometheus), enter the URL, and set Authentication if the endpoint needs it. Select Test Connection, and when it reports
Connection successful!, select Save Data Source.The Free plan includes one data source. Connect your data lists every type, what each one enables, and how to connect AWS with a read-only role.5
Send a test alert
For a first smoke test you can skip the previous two steps and send the alert right after the webhook step. The investigation then says it had little data to work with. Send a firing alert in Grafana’s format straight to your webhook. Replace A successful call answers
<token> with the token from your endpoint URL:{"status":"ok","processed":1,"total":1}.6
Open your first investigation
In the sidebar, select Alerts. The test alert appears under Active Alerts. Investigations start automatically when an alert fires, so select Investigations in the sidebar and open the entry named after your alert. A status of Pending or Running changes to Completed when the agent finishes.If no investigation exists for the alert, open the alert from the Alerts list and select Start Investigation. You can tick Deep investigation first to run extra checks on metrics, logs, changes and traces in parallel, which uses more AI tokens.
What you see
The Investigation Details page shows a Summary, the Root Cause, Recommendations, the Findings with their evidence, and an Investigation Timeline listing every query the agent ran. The buttons at the top let you Stop a running investigation, Retry or Expand a finished one, and generate a Report (technical, management or customer). A test alert about a service that does not exist in your data will produce a thin result. The agent can only report what your data source returns, so point a real alert rule at the webhook to see a full investigation.The dashboard shows a Finish setting up card until everything is connected. It lists the next
connections to make, including Slack and the GitHub App, and what each one adds. The Free plan
includes 10 AI investigations and 100 alerts per month.
If something does not show up
Related
- Run and read an investigation: how to read the result of an investigation.
- Work incidents from Slack: get alerts and results in a Slack channel.
- Set up on-call: page the right person when an alert needs a human.