Zuvio Atlas

Root cause, one click, at 3am

One click turns six hours of metrics, processes, logs and firing alerts into a named root cause — then keep asking follow-up questions in the same thread, with your server's live state already in context.

The question you actually have at 3am

An alert tells you a number crossed a line. It does not tell you why, and at 3am the gap between those two things is thirty minutes of squinting at graphs. Atlas hands the graphs to a model along with the processes, the logs and the other alerts, and gives you back an explanation you can question.

servers / web-01 / ai assistant
AIAI Assistant — web-01

Ask anything — I can see metrics, disk usage, logs, processes and alerts.

Diagnose this serverWhat's taking up disk space?What are the top memory consumers?What's writing the most to disk?Why is CPU high right now?
Ask a question about this server…send

One click for the full diagnosis — or ask in your own words. Either way the panel already holds the server’s live state.

What it is given to work with

The quality of the answer is entirely a function of the context, so the request carries the same evidence a competent engineer would open first:

  • Six hours of CPU, memory, disk, load and network for the affected host
  • The top processes by CPU across that window, not just this instant
  • Recent log events from auth, nginx, syslog, PostgreSQL and MySQL
  • Every alert currently firing on the server, and what each one is measuring
  • The services discovered on the host, so the model knows what is meant to be running
  • Host facts — OS, core count, memory size — so a number can be read as a proportion

Answers shaped for acting on

The answer streams back as an ordinary message, and it is asked to earn every sentence from the data rather than hedge:

  • What changed and when, in plain language, with the actual numbers and timestamps.
  • The affected component, named as a real process, path or service rather than a metric.
  • What to do next, ordered so the first step is the one to try, with the specific command where there is one.
  • What it could not tell — where the context is thin, it says so instead of inventing a cause.

It is a hypothesis, not an oracle

A diagnosis is a very fast first hypothesis from an engineer who has read everything — worth starting from, worth checking. Which is why it arrives as a message you can push back on rather than a verdict. It never takes an action on your infrastructure.

Then keep asking

A diagnosis is one answer. The questions that follow it — would restarting the service reclaim the space, is this the same thing that happened last Tuesday, what should I check first — are where the time actually goes. Because the diagnosis is just the first message in the thread, you ask those in the same place, with the server’s live state already in context.

servers / web-01 / ai assistant
AIAI Assistant — web-01

Diagnose this server — what looks wrong right now, and what should I do about it?

AI

Disk on / crossed 94% at 02:51. Growth traces to postgresql query logging — 34 GB written to /var/log/postgresql since a config change six days ago. Request volume is flat, so this is not traffic.

  1. Rotate and compress /var/log/postgresql — 34 GB reclaimable
  2. Raise log_min_duration_statement; at 200ms it is logging most queries
  3. Add a disk alert at 80% so this surfaces before it is critical

Will restarting postgres reclaim it?

AI
Ask a question about this server…send

It already knows the vitals, the processes, the logs and what is firing.

You never paste a metric or describe your setup. Each message carries the latest vitals, the top processes, six hours of trend, the last five alert transitions, recent log lines and every rule currently firing — so a question as vague as “what changed?” has enough to work with.

Same guardrails

The assistant reads. It does not restart services, edit config, or run commands on your hosts — it tells you what it would do and leaves the doing to you.

Your choice of model, or your own key

Atlas is not tied to one vendor. Diagnosis runs on Claude, GPT-4o, Gemini or Llama, and you can bring your own API key for any of them.

Bringing your own key bypasses the platform quota entirely — useful if you diagnose often, if your organisation already has a negotiated agreement, or if you need the traffic to sit inside a vendor relationship you already have a data processing agreement with.

Beyond a single diagnosis

The same context powers three other things:

  • Runbooks. A generated, editable procedure for a recurring alert, so the second occurrence is somebody following steps rather than re-deriving them.
  • Incident narratives.A written timeline of what happened and when, drafted from the incident’s own record — the part of a post-mortem nobody wants to write.
  • Dependency mapping. Inferred relationships between hosts and services, so a cascade can be read as one cause rather than nine separate alerts.

Why not just paste it into ChatGPT yourself

Because the hard part was never the model — it is the context. Doing it by hand means copying six hours of metrics, the top processes, the firing alerts and the right log lines into a chat window, correctly, every time, at 3am. Atlas already holds all of it; you click once. You never paste a metric or describe your setup, and the answer is grounded in this server’s live state rather than your description of it.

Explore more

14-day free Pro trial — no credit card required

Everything monitored. One bill.

Full Pro access for 14 daysNo credit card requiredDrops to the free plan if you don't upgrade