System monitoring

Sometimes you want more than a one-off "what's the temperature?" — you want to watch a machine over time: the CPU climbing under a build, the GPU heating up, a disk filling. That is what the monitoring stack is for. It records CPU, memory, disk, network (per interface, so Ethernet and Wifi are separate), temperatures, and NVIDIA GPU as time series and draws them on a ready-made Grafana dashboard.

You probably already have it. It is a standard part of LeSysBot, not an add-on: lesysbot setup — the wizard the installer runs — copies the stack into ~/.lesysbot/dashboard, asks which Grafana login to use, and starts it. A normal install leaves Grafana on http://localhost:3000, and lesysbot links to it from the status screen.

Everything it exposes binds to 127.0.0.1 only — nothing appears on your LAN — and none of it needs sudo or admin rights. It runs as its own processes beside LeSysBot rather than inside it: containers on Linux and Windows, Homebrew services on macOS.

What the installer did#

The wizard asks two things and handles the rest:

• How to start it. On Linux, if Docker is ready it offers to bring the stack up right there. If Docker isn't installed or the daemon isn't running, it prints the exact commands to fix that — no sudo prompts mid-install — and you can start the stack later.
• Which Grafana login LeSysBot should use. Defaults to admin / admin; the password is typed masked. It is saved to ~/.lesysbot/grafana.env (readable only by you) and applied to the bundled Grafana's first boot, which is how the status screen and the share-dashboard tool authenticate later.

Neither macOS nor Windows requires Docker Desktop. On macOS the installer uses Homebrew and does the whole thing for you — Grafana, Prometheus and the host exporter, running under brew services. It asks one extra question there: whether to install a small helper for CPU/GPU die temperature, defaulting to no (see Temperatures, honestly). On Windows it points you at a native Grafana download and explains how to connect it, mentioning the one-command Docker stack only if Docker happens to be running. LeSysBot finds Grafana on localhost:3000 either way.

Set LESYSBOT_SKIP_DASHBOARD before running setup to skip the whole step — useful for unattended installs.

Starting and stopping it by hand#

From ~/.lesysbot/dashboard (or the dashboard/ folder of a checkout):

./scripts/install-macos.sh   # macOS — Homebrew, no Docker needed
./scripts/install-macos.sh down

./scripts/start.sh           # Linux — up
./scripts/start.sh down      # stop it again

.\scripts\start.ps1          # Windows (PowerShell)
.\scripts\start.ps1 down

One command, either way: each script inspects the machine — chip, sensors, GPU — and does the right thing. Then open http://localhost:3000. The dashboard is in the LeSysBot folder.

macOS has its own script because it doesn't need Docker at all: it installs Grafana, Prometheus and the host exporter with Homebrew and runs them under brew services, so they come back after a reboot. On Linux and Windows the bundled stack needs Docker with Compose v2:

docker compose version      # should print "v2.x"; if not, install Docker first

Install it if missing — Docker Desktop on Windows or macOS, or Docker Engine on Linux. Prometheus, Grafana, and the metric exporters all download and configure themselves the first time you start the stack — there is nothing else to install or wire up.

What you get#

A single System Overview dashboard, titled for the machine it was built for — Linux, macOS (Apple Silicon), Windows. It shows:

• CPU — busy percentage, per-mode usage, load average, core count
• Memory — used, cached, available, and swap
• Disk — space used per filesystem, read/write throughput
• Network — receive and transmit per interface; the interface name tells Ethernet from Wifi
• Temperatures — whichever sensors your machine actually has (see below)
• GPU — NVIDIA utilization, memory, temperature and power draw, or Apple GPU utilization and memory on a Mac

It is built for your machine#

An empty panel and a broken panel look identical. So rather than shipping one dashboard with every panel for every platform and letting the misses render blank, each start script checks what the host can actually report and leaves out the rest. After that, an empty panel means something is genuinely wrong — which is the whole point.

What gets checked:

GPU metrics key on nvidia-smi, never on the card: the exporter works by shelling out to it, so a GPU without a driver can't be read by anything. When the hardware is there but the tool isn't, the script says so and names what to install rather than quietly dropping the row.

Each script also prints optional add-ons at the end — things that would fill in more panels, like the exact modprobe for a missing Linux sensor driver, or a temperature helper on macOS. They are suggestions, never errors, and never run for you.

Some things still decide themselves at runtime, as before: Linux runs everything on the host network bound to localhost (so it reads the real network interfaces, and sidesteps the firewall rule that blocks a container from reaching the host), and picks a containerised GPU exporter when the NVIDIA container toolkit is present or a native one when it isn't. Windows runs Prometheus and Grafana in Docker Desktop with the host exporter native, reached over host.docker.internal, because Docker Desktop's VM can't see the real host.

Temperatures, honestly#

Coverage follows what each OS exposes without sudo, and that varies a lot:

• Linux is the richest — CPU package and per-core, NVMe/SATA disks, chipset and Wifi radio, AMD GPU. If a driver simply isn't loaded, the start script prints the modprobe that would fix it.
• macOS gives you battery temperature and full Apple GPU utilization and memory out of the box. CPU and GPU die temperature need a small helper — Apple publishes those only through a private framework or root-only powermetrics, and LeSysBot never uses sudo. The macOS installer offers to install one (defaulting to no); or run brew install vladkens/tap/macmon yourself and the tiles fill within 15 seconds.
• Windows shows ACPI thermal zones when the firmware provides them. There is no per-component CPU or disk sensor on Windows without a tool like LibreHardwareMonitor.

Keeping it current#

Re-running setup (lesysbot setup, or the installer) refreshes the stack's scripts and dashboards in ~/.lesysbot/dashboard, while never touching the two things you own: .env (ports and Grafana login) and prometheus/ (any scrape targets you added). Re-run your OS's start script afterwards so the dashboard is rebuilt for the current hardware.

Sharing a snapshot from chat#

You can hand the dashboard to someone with a text message. The bundled share-dashboard tool turns "share me the dashboard" into a public link: it publishes a point-in-time snapshot — the current graphs baked in as data — through Grafana to its public snapshot server, so the recipient sees a live-looking dashboard without any access to your machine or your Grafana.

You:  share me the dashboard for a day
Bot:  📊 Dashboard shared — expires in 1d:
      https://snapshots.raintank.io/dashboard/snapshot/…

You choose how long the link lives (an hour up to a month, or never), you can list everything you've shared, and you can delete any of them again — "list my shared dashboards", "delete snapshot 1". Since each link is a public copy of your metrics, share deliberately and delete what you no longer need. Two things to expect: a link you just deleted may keep loading for up to an hour while the snapshot server clears its cache, and a snapshot you publish from Grafana's own browser button (rather than through the bot) can't be deleted by the bot — so share through the bot to keep it removable.

The tool finds Grafana by itself — it checks the port your stack actually uses, then the usual 3000/3001, and verifies that Grafana really answers there. So a stack that landed on 3001 because something else owned 3000 still works. Set LESYSBOT_GRAFANA_URL only if Grafana runs somewhere unusual; even then it is verified before use.

Where it sits with the rest of LeSysBot#

The security boundaries are unchanged — this stack is a separate process family, not a listener bolted onto LeSysBot. Because it binds only to 127.0.0.1, the exposure is the same as anything else you run locally. If you deliberately publish Grafana, put it behind a reverse proxy with TLS rather than moving the bind off loopback. And change the Grafana password from admin before you do — the wizard is the easy place to set it, or GRAFANA_ADMIN_PASSWORD in ~/.lesysbot/dashboard/.env.

The full reference — every port, the .env overrides, how the dashboards are generated, and a troubleshooting list — lives in dashboard/README.md in the core repo.