Terminal - LumiOps

Every device you run.
One governed shell.

Hypervisors, GPU nodes, switches, PDUs, databases and services across every
site — reachable from one list, with the infrastructure context already on
screen and Lumi proposing the next command. No jump box hunt, no
spreadsheet of IPs, no shared root password.

>_Terminal16

⌕

Site

Type

>_api-prod-7 — Hypervisors
no IP · ssh:22 (default)✎

>_checkout-api — Service
no IP · ssh:22 (default)✎

>_db-main-1 — Hypervisors
no IP · ssh:22 (default)✎

>_edge-gateway — Load Balancers
no IP · ssh:22 (default)✎

>_eval-mistral-large — Job
no IP · ssh:22 (default)✎

>_h100-node-1 — Hypervisors
no IP · ssh:22 (default)✎

>_h100-node-1 · GPU3 — Gpu
no IP · ssh:22 (default)✎

>_h100-node-2 — Hypervisors
no IP · ssh:22 (default)✎

>_ib-spine-1 — Switches
no IP · ssh:22 (default)✎

>_lb-tor-1 — Switches
no IP · ssh:22 (default)✎

>_pdu-A1 — PDUs
no IP · ssh:22 (default)✎

>_postgres-main — Database
no IP · ssh:22 (default)✎

>_payments — Service
no IP · ssh:22 (default)✎

>_train-llama-70b — Job
no IP · ssh:22 (default)✎

>_maestro-deploy-role — Iam Role
no IP · ssh:22 (default)✎

>_legacy-circleci-role — Iam Role
no IP · ssh:22 (default)✎

checkout-api×h100-node-1×lb-tor-1×postgres-main×

postgres-main · site_us_east_1 · Databaserecorded

postgres 15.61,392 qpssite site_us_east_1

› resolving postgres-main in site_us_east_1…

› issuing just-in-time credential · expires 60m

› session recording started

✓ connected · RDS

ops@postgres-main:~$ psql -c "select count(*) from pg_stat_activity"

count

-------

187

psql -c "select count(*) from pg_stat_activity"uptimedf -hcontextlumi ask "is this host healthy?"

ops@postgres-main:~$

Click a device · open up to 4 sessions as tabs · type commands · ↑ for history

REACHES EVERYTHING DISCOVERED IN YOUR ESTATE

THE PROBLEM

Getting a shell takes longer than using it.

At 2am the fix is usually one command. The ten minutes before it are spent working out where the box is, how to reach it, and which credential still works. Race the two and see

02:14

Which host is it on?

Search the CMDB, then Slack, then the wiki page from 2023.

02:18

What’s the IP now?

It moved during the last migration. Nobody updated the sheet.

02:21

Which jump box, which key?

The bastion for dc-west, a key in someone’s home directory.

TIME TO FIRST COMMAND · checkout-api

The old way10:14

✓search CMDB, Slack and the wiki2:10

✓find the current IP — it moved2:40

✕ssh via bastion-dcw → no route to host3:30

✕ssh via bastion-use1 → permission denied4:50

✓ask #infra which key is current9:40

✓connected — with no context10:14

With Lumi Terminal0:02

✓click checkout-api0:00

✓just-in-time credential issued0:01

✓connected · context loaded · recorded0:02

Lumi reached a shell 307× faster — with context loaded and the session already on the record.

AI + TERMINAL WORKFLOW

You bring the question. Lumi brings the command.

Ask in plain language inside the session. Lumi reads the host, proposes the exact command, tells you whether it’s safe before it runs, and explains the output. Pick a question and press the buttons yourself.

1

Proposed, never assumed

Every command is shown before it executes. You run it, edit it, or skip it.

2

Classified before it runs

Read-only goes straight through. Anything that writes is flagged and gated.

3

Output, interpreted

The line that matters, and what it means for this incident.

4

Hands off to a Coworker

When it’s bigger than one host, the SRE Orchestrator takes it wider.

AI + Terminal Workflow

checkout-api· site_us_east_1 · Servicerecorded

?why does this pod keep restarting?

LLUMI · PROPOSED

Four OOMKills in the last hour on this deployment. Let me check memory usage against the limit.

kubectl top pod -n payments -l app=checkout-apiread-only · safeRunEdit

ops@checkout-api:~$ kubectl top pod -n payments -l app=checkout-api

NAME CPU MEMORY

checkout-api-7f9c-x2k 412m 251Mi

checkout-api-7f9c-q8m 398m 256Mi

LLUMI · EXPLAINED

Both pods are pinned at the 256Mi limit. It started with v2.14.3 — the pricing-engine integration loads the full catalog into memory. Raising the limit is the fast fix; paginating the fetch is the real one.

kubectl set resources deploy/checkout-api --limits=memory=512Miwrites · prod · needs approvalRequest approvalEdit

› approval requested · service owner notified · SRE Orchestrator informed

Ssara.lee approved · recorded

✓ deployment/checkout-api resources updated

✓ rollout complete · 0 OOMKills in 5m

>_h100-node-1HYPERVISORSsite_dc_west⚠ 1 issue

GPU UTIL · LIVE80.8%

TEMP82°C

ECC 24H14

MODELPowerEdge XE9680

GPUS8× H100 SXM5 80GB

RUNNINGtrain-llama-70b

CABINET · PDUrack-A1 · pdu-A1

OPEN INVESTIGATIONGPU thermal throttling

LAST CHANGEBIOS 2.4.1 · 12d ago

A REBOOT WOULD TOUCHtrain-llama-70b · 16 GPUml-platform serviceGPU0–GPU7rack-A1 power budget

Loaded before your first keystroke — from the estate model, telemetry and investigations.

INFRASTRUCTURE CONTEXT

Know what you’re logged into before you type.

Every session opens with the device’s facts, live metrics, what’s running on it, and whether there’s already an open investigation. You stop being the engineer who didn’t know the node was mid-training-run.

1

Facts from the estate model

Model, firmware, cabinet, power — normalised from every discovery source.

2

Live health, not a snapshot

Utilization, temperature and error counts streaming while you work.

3

What a reboot would touch

The jobs, services and neighbours that depend on this box.

MULTI-ENVIRONMENT VISIBILITY

Datacenter and cloud in the same
list.

One session plane, three ways in. Click an environment to trace how the connection
actually reaches it.

site_us_east_1 · via cloud API — scoped IAM credentialsno inbound firewall holes on any site

site_dc_west

Your datacenter

Hypervisors, GPU nodes, switches and PDUs reached through the site collector — outbound only.

h100-node-1 · ib-spine-1
pdu-A1 · db-main-1

site_us_east_1

Your cloud

Services, load balancers, databases and IAM roles through the provider API with scoped credentials.

checkout-api · edge-gateway
postgres-main · payments

via CERNE

Isolated sites

Boundary-restricted environments brokered through the control plane on an audited path.

relayed sessions
no direct exposure

ONE SEARCH · EVERY SITE ⌕ payments 1 match · 1 site

>_payments— Servicesite_us_east_1

OPERATIONAL ACCESS

Root access, without a shared root
password.

Every command is classified and checked against your role and policy before it
touches the box. Try it — pick a role, then a command.

1 · YOUR ROLE

vieweroperatoradmin

2 · TRY A COMMAND ON checkout-api · prod

kubectl get pods -n paymentsread-only

tail -f /var/log/checkout.logread-only

kubectl scale deploy/checkout-api --replicas=8write

systemctl restart nginxwrite

kubectl delete ns paymentsdestructive

rm -rf /var/lib/dockerdestructive

policy engine · evaluationheld for approval

1parsetokens ok

2classifywrite · service

3RBACoperator · allowed

4guardrail policyprod write → approval

5auditrequest logged

Held for approvalWrites on production need the service owner. Request sent; runs the moment it's approved.

Just-in-time credentials

Issued when the session opens, revoked when it closes. Nothing lives in a home directory.

RBAC per device

Your role decides which hosts you can reach and whether you get a read-only shell.

Session recording

Every session recorded and replayable — the answer to “who ran that?”

Gated destructive commands

Writes on production pause for approval, under the same guardrails as agents.

SESSION REPLAY

“Who ran that?” now has an answer.

Every session is recorded against a named person and replayable keystroke by
keystroke. Scrub to any moment — amber markers are the commands that changed
something, with who approved them.

replay · s_4f21· priya@acme.com · db-main-1 · 02:31 UTCrecorded

› session opened · priya@acme.com · just-in-time credential · expires 60m

priya@db-main-1:~$ uptime

02:31:06 up 42 days, load average: 7.82, 6.10, 4.02

priya@db-main-1:~$ iostat -x 1 1 | grep nvme

nvme1n1 r/s 812 w/s 2,104 util 97.4%

priya@db-main-1:~$ psql -c "select pid,state,wait_event from pg_stat_activity where state='active'"

pid | state | wait_event

88213 | active | DataFileRead

88240 | active | DataFileRead

LUMI · SUGGESTEDAutovacuum on orders is running during peak and saturating IO. Throttle its cost delay rather than killing it.

priya@db-main-1:~$ psql -c "alter table orders set (autovacuum_vacuum_cost_delay = 20)"

◷ writes to production · approval requested from service owner

✓ approved by sara.lee · 02:33:12

ALTER TABLE

priya@db-main-1:~$ iostat -x 1 1 | grep nvme

nvme1n1 util 41.8%

priya@db-main-1:~$ systemctl reload pgbouncer

✓ approved by sara.lee

priya@db-main-1:~$ exit

› session closed · 4m12s · 7 commands · 2 changes approved · archived

04:12 / 04:12

read-only commandchange · approveddrag the bar to scrub

BY THE NUMBERS

What changes the day you switch
it on.

0 s

from click to an authenticated, recorded shell

shared SSH keys left lying in home directories

0 %

of sessions recorded and replayable by name

0 types

of device, across every site, in one list

CAPABILITIES

What the Terminal gives you.

Unified device access

Every host, hypervisor, GPU, switch, PDU and service — one click from a shell.

🌐

Multi-environment visibility

Datacenter, cloud and isolated sites in one list, one login.

⌕

Smart search & filtering

Find by name, type or site in a keystroke across every device type.

◈

Infrastructure context

Facts, live health, dependencies and open investigations on every session.

🔑

Operational access

JIT credentials, RBAC, command-level policy, recording and replay.

✦

AI + terminal workflow

Lumi proposes, classifies and explains commands — and hands off when it’s bigger.

WHY IT MATTERS

The shell is where incidents actually
get fixed.

Every dashboard ends with someone logging into a box. Make that step instant,
informed and accountable, and the whole incident gets shorter.

TIME TO FIRST COMMAND

10 minutes

1 click

From device list straight to a live, authenticated session.

SHARED CREDENTIALS

in home dirs

collapsed

Just-in-time access replaces keys that outlive their owners.

“WHO RAN THAT?”

unknowable

replayable

Every session recorded against a named person.

WRONG-BOX MISTAKES

common

prevented

Context and issue flags on screen before you type.

Get started

Stop hunting for the box.

Connect one datacenter site and one cloud account. Every
discovered device is one click from a governed shell.

1

2

3

4

1

2

3

Your datacenter

Your cloud

Isolated sites

Build

Orchestrate

Cowork

Resolve

About Us

About Us

Product

Coworker

Orchestrator

Product

Orchestrator

Coworker

Build

Platforms

Estate Models

Agents

Observability

Service Desk