lhm-streamdeck — macOS and DGX Spark fork

A fork of moeilijk/lhm-streamdeck that runs the plugin natively on macOS and monitors NVIDIA DGX Spark (GB10, aarch64) hosts alongside ordinary Linux servers.

Upstream's own documentation is preserved at docs/README.upstream.md and still applies to everything this fork does not change.

Why this fork exists

1. macOS support. LibreHardwareMonitor and HWiNFO are Windows applications, so their Stream Deck plugins are Windows-only. Upstream's manifest declared "OS": [{"Platform": "windows"}] and shipped lhm.exe, so it would not install on a Mac — even though the Go source already cross-compiled to darwin/arm64 untouched, because the Windows-specific code was already isolated behind build tags for the Linux/OpenDeck build.

The port itself is small. macOS reads sensors from lhm-companion over plain HTTP exactly as Linux does; only Windows needs the .NET lhm-bridge subprocess. The one genuine platform difference is syscall.SysProcAttr{Pdeathsig:...}, which does not exist on darwin, so the child-process handling is split per-OS.

2. NVIDIA DGX Spark support. Upstream's lhm-companion publishes linux_amd64 binaries only. A DGX Spark is aarch64, so this fork builds the agent for linux/arm64 too and ships a single self-extracting installer that carries both architectures.

3. Monitoring more than one machine at once. Running tiles against several hosts surfaced a set of cross-host bugs, fixed here (see below).

4. SparkDash LLM performance. Hardware tiles show GPU load and power, not how fast a model is actually generating. This fork can poll a sparkDash dashboard as another source profile and bind live decode/prefill tok/s, KV cache and queue health to keys.

Architecture

      Mac (Stream Deck)                       Remote hosts
 ┌──────────────────────────────┐
 │ Elgato Stream Deck           │
 │  └ com.moeilijk.lhm.sdPlugin │
 │     ├ lhm         (universal)│──HTTP──▶ linux-srv-1:8085  lhm-companion (amd64)
 │     └ lhm-companion  ◀─spawn─┤──HTTP──▶ linux-srv-2:8085  lhm-companion (amd64)
 │        (macOS, :8085)        │──HTTP──▶ dgx-spark-1:8085  lhm-companion (arm64)
 └──────────────────────────────┘──HTTP──▶ dgx-spark-2:8085  lhm-companion (arm64)
                                  HTTP/HTTPS──▶ sparkDash :5555 (or :443)
                                                LLM tok/s, KV cache, queue

Hardware hosts still serve Libre Hardware Monitor /data.json, so GPU/CPU tiles treat every machine the same. SparkDash is an extra source kind on top of that: one profile per dashboard unit, polling model-performance JSON rather than sensors.

What this fork adds

macOS plugin	Universal (arm64 + x86_64) binary, mac added to the manifest
mac-companion/	New macOS metrics daemon serving Libre Hardware Monitor-format /data.json. Auto-spawned by the plugin; no separate install, no sudo
arm64 agent builds	lhm-companion cross-compiled for aarch64 so DGX Sparks are supported
Self-extracting installer	One file carrying both Linux architectures, with SELinux, firewalld/ufw and old-systemd handling plus a --diagnose mode
Lab tile style	Optional per-key renderer with a host badge — brand icon, accent rail and host name — so a key says which machine it is watching
SparkDash LLM source	Polls a sparkDash dashboard for live model performance (decode/prefill tok/s, KV cache, queue) and binds those readings to tiles
pkg/vecicon	Minimal SVG path rasterizer (including elliptical arcs) so icons scale to any size instead of shipping per-size bitmaps

Bug fixes that also affect Windows

These are not macOS-specific; they were found while running against several hosts and apply to upstream on every platform.

• Property inspectors wiped every source profile. Stream Deck's setGlobalSettings replaces the stored object rather than merging it. The inspectors wrote partial payloads ({ pollInterval }, { derivedPresets }), so each write dropped sourceProfiles; the plugin then saw an empty list, synthesised a lone "Default" and persisted it. Changing the poll interval was enough to lose every configured host.
• Switching a tile's source kept the old sensor. Sensor IDs are namespaced per source (/apple/cpu/0 on a Mac, /cpu on Linux), so the carried-over ID resolved to nothing — and the failed lookup was reported as "Libre Hardware Monitor Unavailable" about a host that was answering in under 100 ms.
• Category filter kept the current sensor selected, so filtering to GPU or Memory still showed the previously chosen CPU readings.
• Unknown profile IDs resolved to localhost, because an empty host becomes 127.0.0.1 — a tile referencing a deleted profile quietly displayed local readings under another machine's name.

Install on macOS

Sideloading a built plugin

Stream Deck plugins from outside the Marketplace are sideloaded. Either:

Double-click. Open com.moeilijk.lhm.streamDeckPlugin; Stream Deck installs it and prompts about the unknown developer. This fork is not signed or notarised, so if macOS quarantines the download, clear it first:

xattr -cr ~/Downloads/com.moeilijk.lhm.streamDeckPlugin

Or copy the folder manually, which avoids Gatekeeper entirely:

cp -R com.moeilijk.lhm.sdPlugin \
  ~/Library/Application\ Support/com.elgato.StreamDeck/Plugins/
xattr -cr ~/Library/Application\ Support/com.elgato.StreamDeck/Plugins/com.moeilijk.lhm.sdPlugin
# then restart Stream Deck

Building it yourself

Binaries you compile locally carry no quarantine attribute, so this is the smoothest route. Requires Go 1.25+ (brew install go):

./scripts/macos/build-plugin.sh     # universal plugin + bundled mac companion
./scripts/macos/install-plugin.sh   # install and restart Stream Deck
./scripts/macos/package-plugin.sh   # optional: .streamDeckPlugin for sharing

The plugin auto-spawns the bundled companion on 127.0.0.1:8085 for local Mac metrics.

Install on Linux servers and DGX Sparks

./scripts/macos/build-agents.sh          # linux amd64 + arm64
./scripts/macos/make-agent-installer.sh  # -> build/agents/install-lhm-companion.sh

Copy that one file to each host — scp, USB, whatever works — and run it there. It needs no network access:

sudo sh install-lhm-companion.sh                 # install and start on :8085
sudo sh install-lhm-companion.sh --open-firewall # also open the port
sudo sh install-lhm-companion.sh --diagnose      # report state, change nothing
sudo sh install-lhm-companion.sh --uninstall

If you can SSH to the hosts from your Mac, scripts/macos/deploy-agent.sh user@host does the copy, install and verification in one step instead.

It detects x86_64 vs aarch64, smoke-tests the binary, writes a systemd unit, and verifies the port is actually listening before exiting. On failure it prints arch, OS, systemd version, SELinux mode and binary label, listening state, firewall and recent journal lines.

It handles three things that commonly break this on RHEL-family hosts:

• SELinux — a binary copied out of /tmp keeps a label systemd refuses to execute (status=203/EXEC); the installer relabels it.
• Old systemd — ProtectSystem=strict needs systemd 232; RHEL/CentOS 7 ships 219 and rejects the unit, so it falls back to full.
• firewalld / ufw — detected and reported with the exact command, or opened with --open-firewall.

Configuring the deck

Add a Libre Hardware Monitor → Settings key first and create one source profile per host (host + port 8085), giving each an icon and accent colour. Every key pointing at that host inherits them. Then each sensor key picks a profile, a sensor and a reading.

To poll LLM model performance from sparkDash, add another source with Kind set to SparkDash — one profile per dashboard unit. Point host and port at the dashboard (default 5555 for SparkDash’s own HTTP listener, 443 if you reach it over HTTPS). Pick a unit, Save, then bind tiles as usual. Live decode/prefill tok/s, KV cache and queue readings show up as an LLM category. Lab-style tiles for token rates use tok/s / prefill/s as the middle label and omit the unit next to the number.

Per key, Tile style chooses between upstream's classic histogram and the Lab style shown below. Middle label overrides the label under the value, and hiding it gives that space back to the graph.

Value colour follows a green/amber/red ramp against the key's configured range, so set Min/Max sensibly — a GPU at 72 °C reads red against 0–100 but green against 30–90.

Known limitations

• No Mac CPU/GPU temperature. Apple Silicon die temperatures come from the SMC, which recent macOS no longer exposes to unprivileged processes (powermetrics: unrecognized sampler: smc). The companion omits the Temperatures section rather than reporting fabricated values. Linux and DGX hosts report temperatures normally.
• DGX Spark GPU fields. Verified against two GB10 units: GPU temperature, load, memory load, power draw and core clock all read normally. Only fan.speed is absent — GB10 has no discrete fan to report — so no Fans section appears. Note the Spark's unified memory means "GPU memory" is not a separate pool from system RAM.
• lhm-companion has no authentication and binds 0.0.0.0. Fine on a trusted VLAN; firewall it or put it behind a VPN otherwise.
• SparkDash has no authentication on its HTTP API either. Same trust model: only reach it on a private network or behind your own reverse proxy.
• Binaries here are unsigned and unnotarised.

Credit and licence

All of the hard work is moeilijk's — this fork adds macOS and aarch64 support on top of it. lhm-streamdeck and lhm-companion are GPL-3.0 and this fork inherits that licence.

The Apple, NVIDIA and Linux marks used as host badges come from simple-icons (CC0-1.0). They remain trademarks of their respective owners and are used only to label which machine a key is monitoring. This project is not affiliated with or endorsed by Elgato, Apple, NVIDIA, moeilijk or Mia'a AI Lab.