System inspector
The System Inspector gives you real-time visibility into your Companion Hub machine — resources, containers, ports, and storage — from the browser.
Opening it
Go to Settings → System in the Hub dashboard. The Hub stack and maintenance controls sit at the top of that tab, and the inspector is below them. On a narrow screen the row of Settings tabs scrolls sideways, so swipe the row to reach System.


The page polls every 5 seconds while it is open and includes a Refresh button for an immediate re-read.
The five sections
Everything the inspector shows sits on one tab and refreshes on the same 5-second poll:
flowchart LR
S["Settings → System<br/>polls every 5 seconds · Refresh re-reads now"]
S --> R["Host vs. container resources<br/>physical host vs. Docker VM · recommended Docker memory · VM wedge badge"]
S --> H["System health<br/>CPU · memory · disk · uptime · Docker version · container count · CPU model"]
S --> C["Docker containers<br/>status · uptime · owning app · System for containers Hub did not start"]
S --> P["Port allocations<br/>host port → container · allocated vs. bound · Untracked toggle"]
S --> D["Storage<br/>app data path · Open app data folder in the desktop app, or Copy path"]

Host vs. container resources
This section separates the resources on your physical host from the resources allocated to the Docker VM. On macOS and Windows, those numbers differ, and the Docker VM allocation limits your apps.
It shows Hub’s recommended Docker memory for this machine, says whether automatic Docker memory tuning is on, and flags a VM wedge detected badge when the Docker VM enters an unhealthy state.
System health
Live figures as stat cards:
- CPU load, as a percentage
- Memory used, as a percentage
- Disk used, as a percentage
- Uptime, with the host platform underneath
Below those: the machine’s hostname, the Docker version, a container count split into running and stopped, and the CPU model.
Docker containers
Every container on the host with its status, image, ports, uptime, and the Hub app it belongs to — containers Hub did not start are labeled System. The image column drops off at narrower widths, and so do ports, uptime, and the app column.
Port allocations
What is bound where. Each row gives the host port, the container it maps to, protocol, label, app, and container name, and distinguishes a port that is allocated but not bound from one that is bound and active.
Ports in use on the machine that Hub did not allocate are listed separately as Untracked behind a toggle — which is how you find the thing that is holding the port your app wanted.
Storage
Shows the host path of your app data folder. Inside the desktop app it also offers an Open app data folder button. In a browser there is no folder to open, so the same place shows a Copy path button that copies the path instead.
Apart from Refresh and the folder and path buttons, the System Inspector is read-only. To start, stop, or restart one app, open its page from Home. To act on every app at once, use Batch actions on Home.
Trust & protocol boundaries
Below those five hardware and container sections, a read-only Trust & Protocol Boundaries table summarizes which optional trust protocols and privacy gates are active or fail-closed on this Hub, reading runtime hints from /api/user-context: did:key (SIOPv2) (CI_HUB_DID_AUTH), KERI OOBI (CI_HUB_KERI_OOBI), Cloud Inference Fallback (CI_HUB_ALLOW_CLOUD_FALLBACK), and always-on Forward-Auth EdDSA. See Trust and identity standards for activation instructions and protocol comparisons.
The Resources page
The inspector shows what is on the machine right now. Resources, in the dashboard’s top navigation, shows how your apps, your AI engines and the Hub use the machine over time. It has four tabs and opens on Performance. (Hub v0.3.11 and earlier show an older single-screen Resource Monitor instead; the tabs arrived after it, in CI-Hub #2451.)
| Tab | What it shows |
|---|---|
| Performance | CPU use over a time range you pick, from the last 30 minutes to the last 48 hours, drawn from a reading the Hub takes every 5 seconds. How memory is used and which apps use the most. GPU memory, including what AI models hold. Your AI engines, and the models in memory now with when each unloads. |
| Apps | Each app you can see, with its CPU and memory, shaded by how much of the machine it uses, and its containers. An app shows Running, Not responding or Stopped. |
| AI pool | The other Hubs in your pool, whether each is connected, and their speed. The models reachable through the pool. Each Hub links to its own Resources page. It matters only when you use Hub Pool. |
| Requests | AI requests per minute over the last 30 minutes, and a log of recent routing decisions, newest first, with tokens in and out, the time to the first reply and how each request ended. |
A Needs your attention strip sits above the tabs when something needs you: an app that isn’t responding, a Hub in your pool that can’t be reached or is failing its health checks, a pool setting that isn’t taking effect, AI requests that found no Hub with the model they asked for, a request close to its time limit, app figures that have stopped updating, or a reading that didn’t load. Most notices have a button that opens the tab, or for a pool setting the Hub Pool settings, that explains them.
The Requests tab counts the requests that entered the pool at this Hub: from its apps, from anything on its network or tailnet, and from clients with an API key, wherever they ran. Apps that call an engine on this Hub directly are not counted, and neither is work other Hubs send here. The Hub keeps the last 200 routing decisions in memory, so a restart clears the log.
If you don’t manage the Hub, Resources shows only the apps you can see on Home. The Hub’s own row shows for everyone. When the Hub runs inside a virtual machine, such as the Docker or WSL2 VM, its CPU figures are the VM’s share of the CPU, not the whole computer’s.
GPU memory per app needs a per-process reading: a host probe file, or
nvidia-smi or rocm-smi run by the Hub. Where neither is available, those
figures are left blank rather than shown as zero. Apple GPUs expose no
per-process memory to any tool, so on a Mac they stay blank.