LabScope — Mission control

Five machines.One scope.

An agentless, read-only control room DK built to watch every box, model and AI call in the lab.

Nothing installed. Nothing written. Never sudo.

LabScope

Overview

Metrics every 15 s · read-only

  1. CH 01

    DGX Spark

    Model host

    GB10 Grace Blackwell · 128 GB

  2. CH 02

    Strix Halo

    Image & video

    Ryzen AI Max+ 395 · 126 TOPS

  3. CH 03

    Intel NUC

    Voice

    Core Ultra 9 285H · 99 TOPS

  4. CH 04

    Mac mini M4 Pro

    Control plane

    12 cores · 38 TOPS

  5. CH 05

    Mac mini M6

    Home of LabScope

    12 cores · 24 GB

How it watches

It looks. It never touches.

One SSH session per poll per host. It reads, then leaves.

The cadence

  1. 15 s

    Metrics, every host

  2. 60 s

    Inventory: models, versions, services

  3. 30 d

    Of full history

  4. 2 y

    Of hourly capacity rollups

  1. 01

    Agentless

    Nothing installed on any machine.

  2. 02

    Read-only

    Nothing written, never sudo.

  3. 03

    Redacted

    Secrets are scrubbed before anything is stored.

  4. 04

    Private

    Prompts and answers are never kept.

Page tour

Nine rooms of the control room.

Five of these are real screenshots of LabScope. Network and Fix list are drawn, because the real ones show what guests should not see. Keep scrolling to walk the floor, one room at a time.

OverviewScreenshot
LabScope visitor view: today’s calls, models, tokens made in the lab, uptime, the five machines and the speed leaderboard

01

Overview

Every server is a live scope channel.

InferenceLive

~54tok/s

Qwen 3.6 35B

~30tok/s

Qwen 3.8 27B

KV cache

Time to first token

02

Inference

Live tokens per second, KV cache and time to first token. Qwen 3.6 35B runs at ~54 tok/s, Qwen 3.8 27B at ~30.

ModelsScreenshot
LabScope: what this lab can do, model counts per capability, lab and cloud

03

Models

A memory fit check for all 21 lab models across 251 GB of GPU and unified memory.

AI gateway storyScreenshot
LabScope gateway story: one real request replayed, from app to key, guardrails, routing, model and answer

04

Gateway story

One real request replayed: app → key → guardrails → routing → model → answer.

AI usageScreenshot
LabScope: when the lab thinks, AI calls by hour and weekday

05

AI usage

A live Sankey of every app flowing into the models it calls, and a map of when the lab is busiest. 30 days: 10,635 calls, 62.4M tokens, 96.7% success.

NetworkNo addresses, ever
  • AI lab
  • Whole network
  • Physical
  • Devices
  • Wi-Fi & smart home
4 internet links Firewall Gateway Cloudflare Outbound tunnels only
Isolated AI segment DGX SparkStrix HaloIntel NUCMac mini M4 ProMac mini M6

06

Network

Five views: AI lab, whole network, physical, devices, Wi-Fi & smart home. No addresses, ever.

CapacityScreenshot
LabScope capacity: where would a new workload go, each machine’s busy-hour memory and CPU after adding it

07

Capacity

Busy-hour load against targets, room left per kind of work, and a placement advisor for new workloads.

BenchmarksEvery model type
  • Chat
  • Speech
  • Image
  • Video

Qwen Image · ComfyUI on Strix Halo

357Images

0Failures

32 sMedian

19 sBest

  • Qwen 3.6 35B~54 tok/s
  • Whisper100% words
  • LTX 2.5 22B~4.7 min / clip

08

Benchmarks

Chat, speech, image and video, measured. Qwen Image: 357 images, 0 failures, median 32 s.

Fix listFrom the latest audit

Now

AI advisor

Next

AI advisor

Later

AI advisor

09

Fix list

Audit findings ranked Now, Next and Later, each with an AI advisor from a local model.

0109

Scroll to move through the rooms

Ask the lab

Ask it. Out loud, if you like.

Ask the lab

Type or speak a question. The lab's fastest local model answers from the live snapshot, with tools for web research, installed versions, usage and the fix list.

You

How busy was the gateway today?

The lab · local model

1,098 AI calls and 6.6 million tokens so far today.

From live snapshot · DGX Spark
  • Web research
  • Installed versions
  • Usage
  • Fix list
Ask about the lab…

Only a scrubbed web query ever leaves the lab.

Voice orb

Tap and talk. Whisper hears it on the Intel NUC, Qwen 3.8 27B answers, and Kokoro speaks each sentence as it's written. If the first word takes more than 3 s, the fast 35B takes over.

WhisperHears · Intel NUC
Qwen 3.8 27BAnswers · DGX Spark
KokoroSpeaks · Intel NUC

First word over 3 s

Qwen 3.6 35BTakes over

Alerts & presenter

It finds you. Then it fills the wall.

Alerts reach your phone through ntfy or Telegram, de-duplicated, with quiet hours respected. Presenter mode puts the whole lab on a TV wall.

LabScopePresenter mode21:36
DGX Spark94%
Strix Halo61%
Intel NUC18%
Mac mini M4 Pro34%
Mac mini M622%
AlertDGX Spark · GPU 94% for 10 min21:04
Presenter mode: the whole lab on a TV wall.

Alerts

  • LabScope

    now

    Tunnel back up

    Mac mini M4 Pro · Telegram

  • LabScope

    21:04

    GPU 94% for 10 min

    DGX Spark · ntfy

  • Quiet hours, held

    → 07:00

    Container healthy again

    Intel NUC

  • ntfy
  • Telegram
  • De-duplicated
  • Quiet hours

LabScope

The lab watches itself.

Every 15 seconds, across five machines, four chip makers and 1,263 TOPS, read-only.