Nova Cloud
Nova Cloud is the managed, zero-setup way to run Nova, your AI platform engineer and the trust layer for AI operations. It is not a chat assistant that only answers questions. Nova works across your stack the way a platform engineer does: it deploys and configures, investigates incidents like an SRE, and operates production day to day. And unlike a read-only assistant, Nova can act, with every action governed by guardrails you control.
You don't host anything: AstroPulse runs the Nova orchestrator engine, the skills, the models, and the control layer. You bring the intent.
The operator model
Nova follows the same loop on every task: investigate, plan, approve, execute, observe. It reasons over live evidence and your platform's accumulated operational knowledge (skills, playbooks, lessons learned from past incidents), proposes a course of action, and changes production only after you have approved. The hard part of production AI was never the model. It was the control layer that decides what is safe to run, and that is the part Nova is built around.
Architecture at a glance
Nova separates deciding from doing. The orchestrator is the control plane: it reasons, plans, and enforces guardrails. A data plane (agentgateway) routes model and tool traffic so every call is policy-checked and observable. And changes are applied through the governed Astro Platform runtime, never by reaching into your systems directly.
Under the hood, Nova is a real orchestrator engine built from distinct layers, each its own concern:
- Interfaces — how you reach Nova and manage your organization, access, and usage.
- Reasoning — where Nova decides what to do: answer a question, investigate an incident, or carry out a plan.
- Skills — the toolbox. Each skill is one capability, added without retraining, whether built-in or bring-your-own.
- Data plane (agentgateway) — safe, governed plumbing to your tools and AI models: routing, credential injection, rate limits, and observability.
- Control layer — the guardrails, and the part that makes acting safe: ask-before-acting, a full audit trail, per-tenant isolation, and policy.
- Memory — conversations, replayable investigation history, and the lessons Nova carries from one incident to the next.
Skills and the data plane are the connective tissue; the reasoning and control layers are the engine. That separation, deciding apart from doing, is what lets Nova act in production safely rather than only chat.
Built for production, at scale
Nova Cloud is a production engine, not a single-session chatbot. The orchestrator is built to run many teams' work and many incidents at once, safely and without losing state:
- Multi-tenant by design. Strict per-organization isolation: one tenant's work, data, and credentials never cross into another's.
- Durable and resumable. Work is persisted as it happens, so long-running investigations and plans survive a process restart and resume, with a replayable history.
- Resilient by default. Calls to models and tools run through the data plane with rate limits, retries with backoff, circuit breakers, and health checks; model requests fall back across providers, so one provider's hiccup doesn't stop the work.
- Scales with capability, not model size. New capabilities are added as skills and selected at runtime. Nova keeps the model's working context bounded and budgeted (tool inputs and results kept intact but cost-controlled), so it stays fast and accurate even with a large toolbox and large outputs, no retraining.
- Streaming and observable. Work streams live and resumably, and every model and tool call is attributed and traceable.
What Nova does
Deploy
Ship applications and infrastructure to production, from generated config to a planned, governed rollout.
"Write Terraform for an S3 artifacts bucket and a CI IAM role."
"Generate a production-ready Helm chart for an nginx web service."
Investigate (AI SRE)
Incident response is the vertical Nova does especially well. It investigates like a staff SRE: forms hypotheses, gathers evidence across your tools, tracks confidence as evidence supports or refutes each theory, converges on a root cause, then proposes a remediation and drafts the postmortem. This is evidence-driven reasoning, not a fixed checklist. Investigations are replayable, and Nova carries lessons from past incidents into new ones, so it gets sharper over time.
"Why is the checkout service returning 5xx errors?"
Operate
Run production day to day: inspect clusters and apps, scale, roll back, and act on findings.
"Roll the payments deployment back to the previous revision."
The trust layer: governed action
The capability a chatbot does not have is the one that matters most: Nova can change production safely. That is what makes Nova the trust layer for AI operations. The model proposes what to do; a policy engine decides what is safe to run, evaluating every action before it executes and returning one of three verdicts: allow, require approval, or block.
- Every action is classified — read, write, destructive, or sensitive. Reads run freely; writes, destructive, and sensitive actions require approval or are blocked by default.
- Policy is layered, and the strictest rule wins — environment tier (production is stricter), your organization's policy, per-user policy, and each skill's own risk profile are combined; the most restrictive verdict applies.
- Sensitive changes pause for a human — approval is explicit and required before anything runs.
- Everything is audited — every decision is recorded: the action, the verdict, who approved it, and the outcome.
- Scoped and isolated — Nova only touches what you allow, and one organization's actions, data, and credentials never cross into another's.
The model expresses intent. The runtime decides what is safe to run.
Extensible by design
Every capability is a skill, not a hardcoded integration. Nova discovers the right skills at runtime and composes them to do the job. Adding a skill adds a capability — no model retraining or redeploy. You can extend Nova with first-party skills (cloud providers, Slack, GitHub, infrastructure tooling) or bring your own MCP server to add any capability in minutes, no code. Your credentials are never exposed to the model.
See Skills and Custom MCP Servers.
Works everywhere you do
The same engine, skills, and guardrails are available across Nova Cloud, astroctl nova (your terminal), and Slack. Start an investigation in chat, approve a change from Slack, script a deploy from the CLI.
Get started
- Go to Open Nova → and start working.
- Click Skills in the sidebar and enable what you need (cloud provider, Slack, GitHub, or your own MCP server).
- Ask Nova to investigate, plan, or deploy. Review and approve anything that changes production.
That's it. We handle the rest.