The console that answers back
An assistant inside your console that reads your real infrastructure, diagnoses it from evidence, and acts on it — with your permissions, and your confirmation.
Not a documentation chatbot bolted onto a dashboard. It reaches the platform through the same public API your own browser uses, authenticated as you — so it can answer questions about your live account, and carry out the work, without ever holding privileges you don't have.
Ask anything about your infrastructure
Spend, servers, networks, invoices. It reads your real account, not a manual.
It diagnoses, it doesn't guess
“Why is my site down?” gets a walked path and real evidence — not a list of things it might be.
It acts, you approve
Every change is proposed first. Deleting something means typing its name.
Ask about your account
18 read capabilities across your resources, spend, and history — answered from live data rather than a manual.
Compute & storage
- Instances
- Volumes
- Snapshots
- Backups
Networking
- Public IPs
- VPCs
- Site-to-site VPN
- Firewalls
Billing
- Billing summary
- Usage breakdown
- Invoices
- Spend anomalies
Support & docs
- Support tickets
- Documentation search
Organisation
- Projects
- Service catalog
Advice
- Right-sizing and savings recommendations
It's barred from guessing
On the two questions that matter most, the assistant gathers facts first and reasons only from evidence.
"Why can't I SSH into this box?"
The assistant walks the real path: instance state, then public address, then firewall rules — and reports what it found. No speculation about what might be wrong.
"Is this server the right size?"
It pulls the actual CPU, memory, network, and disk series over a 1-, 6-, or 24-hour window and reasons from those numbers. It's explicitly barred from asserting utilisation it hasn't measured.
Nothing happens because a model decided it should
Every claim below is enforced in code — not in prompt instructions.
You ask
In plain language, from any screen in the console.
It proposes
A write action is never executed directly. It comes back as a proposal with a plain-English description of the consequence.
You confirm
One click for routine work. For destructive actions, you type the resource's exact name — checked on the server, not in the browser.
It runs, and it's recorded
The action is audited under your own name. The assistant separately logs what it proposed, what was confirmed, and what it refused.
The risk ladder
Confirmation: One click
- Power a server on, off, or reboot
- Take a snapshot
- Recover a deleted server
- Open a support ticket
Confirmation: One click
- Deploy a new server
- Resize a server
- Peer two networks
- Invite a teammate
Confirmation: Type the exact name
- Delete a server
- Permanently expunge a server
What we guarantee
It runs with your permissions, never more. Every change is proposed and confirmed. Every action is logged under your name.
Exactly your permissions — never more
Every action travels through the same public API as your browser, authenticated as you. If you can't do it, neither can the assistant.
A proposal is single-use
A replayed confirmation cannot deploy the same thing twice.
The destructive gate is server-side
Typing a resource's name is verified on the server. A check in the UI is only a suggestion.
Scoped to one account
It cannot see or touch another customer's resources. Ever.
It speaks first
A separate checker — deterministic code, not the language model — reads your account and surfaces cards before you type anything. It reports only what it can evidence, ordered deliberately: security before savings before tidying, because an admin port open to the world costs more than money.
Security first
- Admin ports open to the whole internet
- Spend running well above baseline
- Actions failing or stuck
- Zone capacity running hot
Then savings
- Public IPs paid for but attached to nothing
- Volumes attached to nothing
- Servers sitting idle
- Usage that hasn't rated
Then housekeeping
- Servers stopped and forgotten
- Verification still pending
- Quota requests awaiting a decision
Grounded, not remembered
Product questions are answered from Neon Cloud's own documentation, searched directly — rather than from the model's training memory.
There's no per-customer training. Answers come from live reads of your account and our documents, so what it tells you today reflects your account today.
Honest by design
- Deliberately not autonomous — it proposes and waits. That restraint is the point.
- A focused write surface of ten actions — each proposed, confirmed, and audited.
- Can run strictly read-only where your organisation requires it.
AI Assistant FAQs
Everything you need to know. Can't find what you're looking for? Talk to our team.
Our solutions team typically replies within an hour during business hours.
Talk to our team- No. Every write action is proposed first with a plain-English description of the consequence, and runs only after you confirm it. Routine actions take one click; destructive actions like deleting a server require typing the resource's exact name, verified on the server. It is deliberately not autonomous — it proposes and waits.
- Exactly yours — never more. The assistant reaches the platform through the same public API your own browser uses, authenticated as you. If your account can't perform an action, neither can the assistant, and it can never see or touch another customer's resources.
- Anything about your live account: instances, volumes, snapshots, backups, VPCs, firewalls, public IPs, VPNs, projects, billing summaries, usage breakdowns, invoices, spend anomalies, support tickets, and documentation. Answers come from live reads of your real account data, not from a manual.
- It is barred from guessing. For “Why can't I SSH into this box?” it walks the real path — instance state, then public address, then firewall rules — and reports what it actually found. For sizing questions, it pulls real CPU, memory, network, and disk metrics over a 1-, 6-, or 24-hour window and reasons only from measured numbers.
- Yes, twice over. Every platform action is audited under your own name, exactly as if you had done it in the console. Separately, the assistant keeps its own log of what it proposed, what was confirmed, and what it refused.
- Ten write actions, each requiring your confirmation: deploy, resize, power-control, delete, recover, or expunge a server; take a snapshot; peer two VPCs; invite a teammate; and open a support ticket. It does not create load balancers, Kubernetes clusters, DNS records, or object storage from chat.
- Product questions are answered from Neon Cloud's own documentation, searched directly — not from the model's memory. There is no per-customer training: answers come from live reads of your account and our documents.