A Windows HUD for a Hermes agent on your own server.

How it reaches the server

• Opens an SSH session with the password or key you entered once.
• Finds Hermes, on the host or inside a Docker container, reading /proc/net/tcp inside each running container when it has to.
• Opens a local port forward to wherever it turned out to be, relaying through docker exec for a container.
• Reads the API key from ~/.hermes/.env over the same session, so you never paste a secret you already own.
• Runs a health check, then the HUD comes up.

What the readings cost your server

One round trip a second batches the whole lot into a single command, with a @@ sentinel between sections, so one read fills every gauge.

# CPU, per core
cat /proc/stat;    echo "@@";
cat /proc/meminfo; echo "@@";
cat /proc/loadavg; echo "@@";
cat /proc/uptime;  echo "@@";
cat /proc/net/dev; echo "@@";
df -B1 -P -x tmpfs -x devtmpfs -x squashfs

Processes every 5 seconds, systemd timers every 60, the watched file tree every 3. At that cadence the whole loop stays a handful of kilobytes a second.

Required

• Windows 10 or 11, x64. No account and no administrator rights: it installs per user, into %LOCALAPPDATA%.
• A Linux server you can reach over SSH with a password or a key. Any distribution that exposes /proc, so all of them.
• Hermes installed on that server with its API server enabled and bound to 127.0.0.1.
• About 3 GB of free space. The installer alone is 1.6 GB, and most of that is the speech models.

Not required

• An account, a subscription, or a sign in of any kind.
• Pasting an API key. JARVIS reads it off your server over the SSH session you just authenticated.
• An NVIDIA card. It makes speech about seven times faster and nothing depends on it.
• A microphone. Voice is optional and the app works just as well by keyboard.
• Opening any port to the internet. Hermes stays on loopback and JARVIS reaches it through an SSH forward.

openWakeWord · hey_jarvis3 MB

Listens continuously for one phrase: "Hey JARVIS". A pretrained model for that phrase alone, small enough to run on the CPU on any machine. The detector is deaf while listening and while speaking, so JARVIS cannot trigger itself.

The pretrained openWakeWord models are non-commercial by their own licence. That restriction applies to this component alone, and it is set out in full in the terms.

faster-whisper75 MB to 1.6 GB

Turns what you say into text, on the CPU or on an NVIDIA card. The models are multilingual, which matters here: the English only models cannot do Arabic at all.

The language detectortiny · 75 MB

A second copy of whisper that never transcribes anything. Its only job is to answer which language was spoken. Identifying a language is an easy problem and transcribing speech is not, so the cheap model answers it and the real model is told the answer.

Over eight clips in English and Jordanian Arabic, clean and noisy: 39.2 seconds before the change and 24.4 after, with identical text on seven of the eight. Below 0.5 confidence the detector is ignored in favour of whatever language the speaker was last using.

Piper63 MB × 2

Speech out. Two voices: en_GB-alan-medium for English and ar_JO-kareem-medium for Arabic. The voice follows the script of the text, so Arabic characters mean the Arabic voice, and that stays right even when Hermes answers in a language you did not use. With no Piper voice present, Windows speech takes over, so JARVIS is never mute.

Two values are yours, not ours

• The container name. The Hostinger one-click template appends a random suffix, so it differs on every deployment and changes again if you redeploy. Step 2 prints it.
• The host address. 172.17.0.1 is the default Docker bridge gateway and holds on most machines. If yours differs, ip route | awk '/default/ {print $3}' inside the container prints the right one.

$CONTAINER lives only in the shell you set it in. Close that SSH session and come back later and it is gone, so run step 2 again before anything that uses it.

One thing to weigh first: on the SSH backend everything runs on the host, so the agent can no longer edit its own code inside the container.

Where a command ends up

1. Hermes container
2. SSH
3. root@172.17.0.1
4. VPS host

The installerv1.20.0

One file, installed for a single user into %LOCALAPPDATA%\Programs\JARVIS with no administrator rights. It carries the application and the voice sidecar together, and most of its size is the speech models, which are what let JARVIS listen and answer with no network round trip.

Or build it yourself

flutter build windows --release

# then the installer
ISCC.exe installer/jarvis.iss

The sound cues and the application icon are both generated by scripts in the repository, so no binary asset is checked in that cannot be produced again.