Everyone can diagnose. Almost nobody can safely act.
Nova works on the stack you already run, with every production action scoped, approved, audited, and reversible.
Everyone diagnoses. The control plane for safe action is the moat.
or start from your terminal
$ astroctl novaCurrently onboarding early design partners.
Your team could be here. Become a design partner.
what nova actually does
Investigate, mitigate, remediate, verify, postmortem — one loop, with a gate where the writes begin.
Ask it
Answers about your own clusters and applications, not generic advice — with the evidence and the pages it used.
Tell it to ship
Describe the change and it writes the artifacts, plans the steps, and carries them out through governed skills.
Page it
It runs the whole loop, and can fix what it finds — once a person opens the gate.
Investigate
Forms hypotheses, collects evidence, names findings. Nothing that can write is on the table yet, and a read that would expose customer data still asks.
Mitigate
Remediate
Verify
Postmortem
earned autonomy
Everything Nova touches goes through Astro Platform. Reads that expose nothing run unattended, reads that expose customer data stop where writes stop, and anything past the approved scope is refused and recorded.
4
ran unattended
Health, events and rollout history. Reading them exposes nothing and changes nothing, so nobody is asked.
2
held at the gate
One read and one write. The logs can expose a customer; the rollback changes production. Both wait.
1
refused outright
Out of scope, so no one was asked. A gate that only ever opens is not a gate.
every step of one incident
Read the fleet
ran
Cluster and application health across the org. checkout-api on prod-us-east is the one that is degraded.
Read the rollout history
ran
v2.3.1 replaced v2.3.0 twenty minutes ago. v2.3.0 had been healthy for six days.
Read the application events
ran
The 5xx rate went from a 0.4% baseline to 3.2%, starting at the minute of the rollout.
Read the application health
ran
The new version is up and passing its checks, and still serving 500s. The fault is the release, not the platform.
checkout-api v2.3.1 is serving 500s.
90% confidence, from three independent signals that agree.
Read the application logs
held
A log line can carry a customer’s data, so a read that exposes one is gated exactly like a change. Approved, and it named the checkout path.
Roll checkout-api back to v2.3.0
held
A production write on a tier-one workload, so it waits for an admin or an SRE. Blast radius: one service on prod-us-east. Approved, then executed and watched back to baseline.
Destroy and redeploy prod-us-east
refused
Past the approved blast radius. Denied and written to the audit trail before the approved rollback ran.
prompt to production · preview
Add astro-deploy to Claude Code, Cursor or VS Code once. It runs on your machine, so your source stays there: the repository is read where it sits, a blueprint is resolved, and the artifacts are written with the production defaults most teams meet during an incident.
local → dev → staging → prod
Deploy builds for the entry environment. Promotion moves the image that was already validated rather than building a new one, and an environment will not take a release until the one before it is healthy.
read from your repository
language TypeScript · framework Next.js · port 3000 · stateful no
blueprint resolved
Server-rendered web apps with a backend. Standard infrastructure.
installs ingress · TLS certificates · DNS records
2 nodes · Deployment
written for you
Dockerfile · Helm chart · GitHub Actions
production defaults you did not have to know
The difference between a service that deploys and one that survives its first bad night.
$ claude mcp add astro-deploy -s project -- mcp-server-astro-deploythe real question
Diagnosis is becoming a commodity. Permission to act in production is not.
The model was never the bottleneck. The control layer was. Everyone can diagnose, and everyone will tell you their agent asks permission. The question worth asking is whether the policy sits in the execution path, or beside it.
Read our take: The AI SRE race is running the wrong way
How we designed Nova's investigation engine: lessons from SRE at scale
see it in action
Real recordings of the real product, acting under governed approval.
Nova applies the fix when it can, and holds off when it should, honest about every write.
opening the gate
The runtime stops the change. A person releases it, without leaving the thread they are already in.
Slack
The plan arrives in your channel with its steps and which of them need a decision. Approve or reject it there, or @mention Nova in the thread to ask why first.
See Nova in SlackAstro Platform
Scope, blast radius and rollback target are decided by the runtime rather than asked of the model, because it reads clusters, rollouts, cost and audit as first-party state.
See Astro Platformskills & integrations
Ten skills today, plus any MCP server you bring — each one scoped, policy checked and audited.
Manage Kubernetes clusters, deploy apps, and analyze infrastructure.
Create and manage Helm charts for Kubernetes. Five actions.
Look up provider docs, verified modules, versions and policies.
Calculate service costs and pricing estimates.
Calculate service costs and pricing estimates.
Calculate service costs and pricing estimates.
Manage repositories, pull requests and workflows.
Chat with Nova directly in Slack.
Connect logs, metrics, monitors, incidents and dashboards.
Connect your runbooks, internal docs and postmortems.
Authenticate once and the skill is Nova's to use — every action scoped to what you granted, policy checked, and written to an audit trail.
One-click connect · governed actions · per-tenant isolation · full audit trail
Paste an MCP server URL, add a bearer token or API key, and Nova discovers its tools in seconds — indistinguishable from the skills we ship.
Streamable HTTP · bearer or API key · auto tool discovery · first-class skill
how you run nova
Managed
RecommendedBrowser, terminal and Slack over one managed backend — start in one, continue in another, same session throughout.
Open Nova CloudSelf-hosted
Early accessDocker Compose, with local model execution for when data has to stay inside your boundary.
Open Nova DirectIn your editor
Request previewA hosted MCP endpoint for Claude Code, Cursor and VS Code. Add the config and Nova investigates from where you already work.
Open Nova ConnectA full interactive chat in your terminal. Paste logs, ask questions, and pick a prior session back up where you left it.
$ astroctl nova$ astroctl nova --continue$ astroctl nova --resume$ astroctl nova sessions list --limit 25Nova is the AI platform engineer. Skills connect it to your systems. Astro Platform is the runtime that unlocks governed production execution.