Use cases

Your servers just hired
a night shift.

The patching, the triage, the "can you check why that box is slow" — it's all SSH work, and AI agents are already good at it. What they lacked was a safe way in. Give an agent a scoped portal to your fleet and most of your server chores become one sentence long: the agent does the legwork, your policy guards the edges, and the risky step waits for your click.

Incidents: from alert
to diagnosis, unattended

The 2 a.m. page used to mean a laptop on the duvet and twenty minutes of ssh, grep and swearing. Now the agent gets there first: it sweeps the tier, reads the logs, correlates across boxes, and arrives at a diagnosis with the evidence attached.

The fix that changes state — a restart, a truncate, a config rollback — queues for approval. You wake up to a finished investigation and one button.

  • Reads are instant; writes wait for you (or your policy lets the safe ones through)
  • Works across the whole tier in one conversation
  • The entire investigation is replayable, evidence and all
agent: ops-bot · scope: api tier (4 endpoints) ● recording
> "api tier is throwing 502s — find out why"

[portal] ops-bot ▸ run_command api-02 systemctl status app
api-02  app.service: activating (auto-restart)
[portal] ops-bot ▸ run_command api-02 journalctl -u app -n 50
api-02  kernel: Out of memory: Killed process 1182 (app)
[portal] ops-bot ▸ run_command api-02 free -m
api-02  Mem: 3908 total, 3841 used — leak since 01:14 deploy
[portal] diagnosis: 01:14 build leaks; rollback queued ▸ your approval 
4 commands, all logged 1 approval pending

Patch 23 servers
with one sentence

Fleet-wide chores are where agents embarrass the old way of working. "Apply the security updates everywhere, reboot what needs it, one box at a time" is an afternoon of careful tab-juggling for a person — and a loop for an agent.

Scope it to the fleet, allow apt in policy, keep reboots behind approval. The agent works through every box, staggers the restarts, skips the one that looks weird and tells you why.

  • Same playbook every time — no box forgotten, no step skipped
  • Rolling by default: it watches each server come back before the next
  • A written report at the end: what changed, what rebooted, what it left alone
agent: patch-bot · scope: all servers (23 endpoints) ● recording
> "security updates everywhere; reboot one at a time"

[portal] patch-bot ▸ run_command web-01 apt upgrade -y
web-01  14 upgraded · needs reboot
[portal] reboot web-01 ▸ approved by policy window
web-01  back in 41s · health check ✓
[portal] patch-bot ▸ run_command web-02 apt upgrade -y
web-02  14 upgraded · needs reboot
[portal] …21 more · db-01 held back: kernel pin — see note 
22/23 patched rolling reboots, health-checked

The chores nobody does
until they bite

Backups that were never test-restored. Certificates that expire on a Saturday. Disks that fill at the worst moment. These fail because they're boring — and boring is exactly what agents are for.

Run them on a schedule with hosted agent runs: your Claude Code, Codex, Gemini or Grok account, launched in a locked-down container we host, whose only reach is your fleet. Every run leaves a transcript, so "did it actually check?" has an answer.

  • Nightly: restore the latest backup into scratch, row-count it, report
  • Weekly: sweep every cert on every box, renew what's inside 30 days
  • Daily: disk hygiene — rotate, prune, truncate, and say what it freed
hosted run · nightly · backup-verify ● recording
> "restore last night's dump into scratch and verify"

[portal] agent ▸ run_command backup-01 ls -t /backups | head -1
backup-01  prod-2026-11-21.dump · 4.2G · 02:00
[portal] agent ▸ run_command db-scratch pg_restore …
db-scratch  restore complete in 6m12s
[portal] agent ▸ run_command db-scratch psql -c 'SELECT…'
db-scratch  41 tables · 18.2M rows · counts within 0.3% of prod
[portal] report: backup verified · transcript saved 
runs nightly, unattended scratch db only — prod is out of scope

A dozen more things
to stop doing yourself

Each of these is a real ask, verbatim. Scope the agent to the right boxes, and say it.

Ship v2.14 to staging, smoke-test it, then prod

Deploy, verify, promote — with the production step held in the approval queue.

Who's hammering the API? Block them

Tails access logs, finds the abuser, proposes the firewall rule for your click.

Set up the new box like the others

Hardens a fresh server to match the fleet: users, firewall, fail2ban, unattended upgrades.

Why is this server slow every day at 4pm?

Correlates cron, I/O and process history across days until it finds the culprit.

Audit every box for users who shouldn't exist

Sweeps /etc/passwd, authorized_keys and sudoers fleet-wide; reports drift.

Migrate these cron jobs to systemd timers

Converts, installs, verifies the next-run times, and removes the old crontab lines.

Check replication lag and failover readiness

Runs the checks you never get to, on a schedule, and only speaks up on drift.

Clean up docker on the build hosts

Prunes images and volumes with a size report — the destructive flags behind approval.

Tighten nginx TLS config everywhere

Updates ciphers and HSTS across the web tier, reloads, and verifies with a probe.

Find what changed before the outage

Diffs package history, config mtimes and logins across the window — then shows its work.

Rotate the app's database password

Generates, updates the vault'd config, restarts in order — each write step approved.

Tell me every morning if anything looks off

A scheduled sweep: disks, RAID, queues, certs, load — silence when all is well.

Autonomy you dial in

Start with everything behind approval. Loosen policy as the agent earns trust — per command pattern, per server. Full autopilot for reads, a click for writes is where most teams land.

Blast radius, contained

Scope means a misbehaving agent can only reach the boxes you attached. Zero keys on the agent, five-minute certs and one-session-one-server isolation do the rest. The security model →

Receipts for everything

Every ask, every command, every output — hash-chained and replayable. When the agent did the work, you can show exactly what the work was.

The chores aren't going
to do themselves. An agent will.

Add a server, deploy an agent, and hand over your first chore in under five minutes.

Get started · $9.95/mo