Configure
Nova works with safe defaults, so you can start without changing anything here. Organization owners and admins change these settings. Everyone else can see them, except Incidents, which only owners and admins see.
Skills Nova reads from
Open Skills in Nova. See Skills.
| Skill | What it adds to an investigation |
|---|---|
| AstroPulse (required) | App status, configuration, Kubernetes and platform events, deploy history, other apps on the same cluster, add-on status and, with approval, app logs. Also the actions fixes are built from: update, roll back, resync or resume an app, and change a node group's size. |
| GitHub | The commit and pull request a deploy was built from. |
| Slack (Enterprise plan) | Starting investigations from a thread, and alert results in a channel. |
| Datadog (Enterprise plan) | Logs, metrics, monitors and traces, when they help tell causes apart. Each person connects their own Datadog keys. |
| Your own MCP servers (Enterprise plan) | Their read tools, when they help tell causes apart. Every call asks a person until an admin trusts the server. |
When a skill that is not connected would have helped, the investigation says what it would have shown for that app and how to connect it.
To connect each skill, see Skills.
Guardrails
Open Settings (the gear at the bottom of the sidebar) and choose Guardrails. The categories that matter most for incidents:
| Category | Default | During an investigation |
|---|---|---|
| Observability & Monitoring | Allowed | Nova reads health, metrics and events without asking. |
| Sensitive Data Access | Requires approval | Nova asks before it reads app logs or security scan results. |
| Changes & Remediation | Requires approval | A fix always waits for approval, whatever this rule says. |
| Scaling & Resizing, Infrastructure Provisioning | Requires approval | A change of this kind always waits for approval, like a fix. |
| Destructive Actions, Secrets & Credentials, Infrastructure Teardown | Blocked | Never run. Listed under Could not check if Nova needed them. |
| Messaging & Notifications | Allowed | Slack messages Nova posts. |
Use Shadow mode to trial a stricter policy, and Emergency stop to halt every write while Nova keeps reading. See Guardrails and Approvals.
Your own MCP servers
A server you connect starts untrusted: every call to one of its tools asks a person first, including reads during an investigation, whatever the server says about the tool.
To trust it, an organization owner or admin opens Skills, opens the server's settings and turns on Trust this server. Nova then uses the server's own labels for which tools read and which write, under your guardrails. The settings show who trusted it and when. Trust never lets a call skip what your guardrails require, and any change to the server's tools, or to its address, removes it.
Who approves, and from where
In Nova, open Settings → Guardrails and find Who approves.
approverRoleslackApprovalsSlack is on Enterprise.
standingLogAccessAppsEvery other app is asked about, and every read is in the audit log.
Only people who may approve can allow a sensitive read. In Slack, a person is matched to their organization account by email.
Your team's conventions
In Settings → Instructions, under Instructions for Nova, write what Nova should know about your team: where services live, who to page, what never to touch (up to 4,000 characters). Nova uses it as context in investigations and postmortems. It never unlocks an action.
Share what Nova learns with the team is on by default, so a cause ruled out in one engineer's investigation helps the next person on call.
Alerts and the Slack channel
An alert-started investigation is a deep investigation (10 units) on every plan. In Settings → Incidents, create the alert token and, on Enterprise, choose the Slack channel for results. See Start an investigation and Alerts.
The same page holds how incidents are worked in Slack, and the time zone Nova writes in:
incidentChannelsEach incident gets its own Slack channel, whatever started it. Off, each incident stays in the thread it started in. Turning it on needs Slack connected, on Enterprise.
incidentChannelPrefixShown when channels are on. How each incident channel's name starts: the prefix, the incident number and the app. The page shows the name the next channel would get. Applies to channels opened from then on.
- and _, up to 20, starting with a letter or digit.incincidentChannelRecogniseShown when channels are on. Name prefixes of channels your team or paging tool already opens for incidents. When someone asks Nova in such a channel, Nova works there instead of opening another.
timeZoneThe zone Nova writes times in for updates it posts to an alert channel, for the postmortem file, and for on-call emails to people who have not chosen their own. The zone is always named.
Limits
Each limit sits with the settings it belongs to. The first three are in Settings → Guardrails, in Advanced under Who approves. The next four are in Settings → Instructions, under Investigation limits. The last five are in Settings → Incidents, under Advanced.
approvalTimeoutMinutesaccessGrantMinuteshealthVerifyMinutesinvestigationMaxStepsinvestigationMaxDurationMinutesinvestigationMaxRetriesconcurrentInvestigationsalertInvestigationsPerHouralertGroupingWindowHoursalertStormThresholdstakeholderUpdateMinutesstalledAfterMinutesSteps and time apply to a standard investigation. A deep one gets twice both, up to 20 steps and 60 minutes. Time spent waiting on a person does not count. See Standard or deep.
When more investigations start than Investigations at once allows, the next one waits its turn and starts when one finishes. None is dropped.
A read that fails for a passing reason (a timeout, a server error, a dropped connection, or no reason given) is tried once more after a short pause. A read the service rate-limits is tried once more after the wait it asks for, and a request the tool rejects as malformed is corrected and sent once more when Nova can correct it. Every retry counts toward Retries per investigation.
While service is impaired, an alert's Slack thread gets a one-line stakeholder update (impact, status, next update) at each change and at least as often as this setting.
Cost
An investigation is charged once, when its work is done: 5 Nova units for a standard investigation and 10 for a deep one, or 1 unit if Nova can't read anything. The card shows the size and price before work starts, and with fewer than 10 units left a deep investigation runs as a standard one. An alert-started investigation is never made smaller: with fewer than 10 units left it is refused. Answers, approved fixes and the postmortem cost no further units. An alert's investigation runs as the admin who created or last rotated the alert token, and is paid from your organization's units. See Nova Units.
If it's refused
Next
- Card reference: every status, every answer, the terminal keys and the API.