Automate air compressor monitoring with a Shelly Pro 3EM, InduQT and Claude

Updated Oct 31, 2026 · InduQT engineering team · Guide, not a customer case study

The short answer

A three-phase energy meter with clamp-on current transformers, such as the Shelly Pro 3EM, tells you most of what matters about an electric air compressor without touching its controller: running, loaded or unloaded (from the current), run hours, starts per hour and energy. It publishes over MQTT, costs about US$130, and fits in the compressor’s supply panel. InduQT then learns the compressor’s normal load pattern and warns when it starts running loaded longer for the same demand, which usually means a leak, a fouled cooler or a failing inlet valve.

  • Loaded vs unloaded hours and starts per hour, without touching the compressor controller
  • A warning when the load pattern drifts (leaks, cooler fouling, valve problems)
  • Energy per shift you can put a price on
Costabout US$230 including the gateway
Time1–2 hours (an electrician fits the clamps)
SkillSome wiring
  1. What to buy
  2. Fit the meter
  3. Turn on MQTT
  4. Set up the MQTT broker (once per plant)
  5. Check the readings with Claude before you connect anything
  6. Wire it to InduQT
  7. What InduQT does with it

What to buy

Shelly Pro 3EM-120Aabout US$130

DIN-rail three-phase energy meter with three 120 A clamp-on CTs; Wi-Fi, Ethernet and MQTT. Maker: Shelly. Maker’s page

Shelly Pro 3EM-400A (for compressors over about 75 kW)check the listing

Same meter with larger clamps. Maker: Shelly. Maker’s page

Raspberry Pi 5 (4 GB) with power supply, case and 32 GB cardabout US$100 for the kit

Runs the MQTT broker and the InduQT bridge. Any always-on PC or small industrial PC works too. Maker: Raspberry Pi. Maker’s page

Prices are typical street prices in US dollars when we wrote this, not quotes; buy from the maker or a local distributor. InduQT does not sell sensors and is not paid by any maker named here.

Safety. Fitting the meter means working inside a live supply panel. Have a qualified electrician isolate and lock out the supply, fit the clamps around the phase conductors (not the earth) with the arrows towards the load, and wire the meter’s voltage inputs through suitable fuses.

Fit the meter

  1. Isolate and lock out the compressor supply.
  2. Mount the Pro 3EM on the DIN rail in the compressor’s supply panel or the distribution board feeding it.
  3. Clamp one CT around each phase conductor to the compressor (A, B, C), arrows pointing to the compressor.
  4. Wire the meter’s voltage inputs and neutral as in Shelly’s wiring diagram, and its power supply.
  5. Restore power. The meter’s web page shows current on each phase when the compressor runs.
If the clamps read negative power, a CT is backwards. Flip it rather than fixing it in software.

Turn on MQTT

  1. Power the Shelly and join it to the plant Wi-Fi (or plug in Ethernet on Pro models) with the Shelly app, or open its own Wi-Fi hotspot and browse to 192.168.33.1.
  2. Open the device’s web page (http://<device-ip>) → Settings → MQTT.
  3. Turn MQTT on. Server: 192.168.1.20:1883 (the Raspberry Pi). User and password: the sensors user.
  4. Client ID and topic prefix: factory/compressor-1. The bridge reads the machine name from the level after factory/.
  5. Turn on “Generic status update over MQTT” (status_ntf), so the device publishes its full status to <prefix>/status/<component>:<id> whenever something changes.
  6. Save and let the device reboot. Within a minute mosquitto_sub on the Pi shows its messages.

The meter then publishes factory/compressor-1/status/em:0 with a_current, b_current, c_current, a_voltage, a_act_power, total_current and total_act_power (the field names are Shelly’s, from its EM component), and factory/compressor-1/status/emdata:0 with energy totals.

Set up the MQTT broker (once per plant)

The sensors publish their readings to a broker: a small program that receives MQTT messages and hands them to whoever subscribes. Mosquitto is free and runs on a Raspberry Pi, an old PC or a small industrial box on the plant network. You need one per plant, not one per machine. If you already have a broker (Node-RED, Home Assistant, an IoT gateway), use it and skip this step.

Install Mosquitto
# Raspberry Pi OS, Debian or Ubuntu
sudo apt update
sudo apt install -y mosquitto mosquitto-clients

# One user for the sensors and one read-only user for testing
sudo mosquitto_passwd -c /etc/mosquitto/passwd sensors
sudo mosquitto_passwd /etc/mosquitto/passwd viewer
Require a password
# /etc/mosquitto/conf.d/factory.conf
listener 1883
allow_anonymous false
password_file /etc/mosquitto/passwd
acl_file /etc/mosquitto/acl
Who can publish and who can only read
# /etc/mosquitto/acl
user sensors
topic readwrite factory/#

user viewer
topic read factory/#
Start it and watch messages arrive
# Mosquitto 2 wants these files readable only by its own user
sudo chown mosquitto: /etc/mosquitto/passwd /etc/mosquitto/acl
sudo chmod 0700 /etc/mosquitto/passwd /etc/mosquitto/acl

sudo systemctl restart mosquitto
sudo systemctl enable mosquitto

# Watch everything that arrives (Ctrl+C to stop)
mosquitto_sub -h localhost -u viewer -P 'viewer-password' -t 'factory/#' -v
Give the Raspberry Pi a fixed IP address (a DHCP reservation on the router) so the sensors can always find it. Every example below uses 192.168.1.20. Keep the broker on the plant network; it does not need to be reachable from the internet.

Check the readings with Claude before you connect anything

Before the data goes anywhere, make sure the sensor is telling the truth. Record two minutes of messages to a file and ask Claude (or ChatGPT) to check them. This catches the usual first-day problems (a clamp on the wrong cable, a probe reading the room instead of the pipe, a sensor that only reports when something changes) while you are still standing at the air compressor.

Record two minutes of messages (Raspberry Pi, Mac or Linux)
# Record 2 minutes of messages to a file you can give Claude or ChatGPT
mosquitto_sub -h 192.168.1.20 -u viewer -P 'viewer-password' \
  -t 'factory/#' -v -W 120 | while read -r line; do
  echo "$(date -u +%Y-%m-%dT%H:%M:%SZ) $line"
done > sensor-sample.txt
The same on Windows
# Windows PowerShell, with Mosquitto installed from mosquitto.org
& "C:\Program Files\mosquitto\mosquitto_sub.exe" -h 192.168.1.20 -u viewer -P viewer-password `
  -t "factory/#" -v -W 120 | ForEach-Object { "$(Get-Date -Format o) $_" } | Out-File sensor-sample.txt
Prompt
Paste into Claude or ChatGPT with sensor-sample.txt attached
I've attached sensor-sample.txt: two minutes of MQTT messages from a sensor I just fitted to the air compressor.
Each line is: time, topic, payload.

Please check it for me:
1. List each topic and every numeric field in it, with the min, max and typical value.
2. Tell me whether the values are physically plausible for the air compressor. Loaded current should be close to the motor nameplate full-load current; unloaded is typically 25–40 % of it; the three phases should be within a few percent of each other. Tell me the loaded and unloaded current you see and how many load/unload cycles happened.
3. Point out gaps longer than 30 seconds, repeated identical values, or fields that are always 0.
4. Say which fields are worth sending to InduQT as signals and which I can ignore.
Answer in a short table, then a list of anything I should fix before I wire it up.

Claude Code or Claude desktop with a terminal can run the mosquitto_sub command for you: ask it to "subscribe to factory/# on 192.168.1.20 as viewer for two minutes and check the readings". Either way the readings stay on your laptop: nothing is stored in InduQT until you wire it up in the next step.

Using a cloud broker (HiveMQ Cloud, EMQX Serverless or your own server with TLS) instead of one on the plant network? Add the free InduQT connector to Claude (https://claude.ai/directory/connectors/induqt) and Claude can listen to the broker for a few seconds itself:

Prompt
Claude with the InduQT connector
Use InduQT to preview my MQTT broker before I connect it:
url mqtts://<your-cluster>.hivemq.cloud:8883, topic factory/#,
username viewer, password <viewer password>. Listen for 10 seconds.
Then tell me which topic level names the machine and which fields are numeric.

Wire it to InduQT

Once the readings look right, send them to InduQT. The simplest way is a small bridge on the same Raspberry Pi: it listens to your broker and forwards every number to InduQT over HTTPS. It works on every InduQT plan, including the free one (up to 3 machines), and nothing on the plant network has to be opened to the internet.

  1. In Claude with the InduQT connector, paste the prompt below. Claude creates the site, the machine and a key that can only send readings to that site. (No Claude? Do the same in the InduQT app: Add machine with the same externalId, then Settings → API keys → a key with the ingest scope, limited to the site.)
  2. Copy the key Claude shows you. It is shown once.
  3. Start the bridge on the Raspberry Pi with that key.
  4. Ask Claude for the latest values, and set the normal bands.
Prompt
1 · Claude creates the machine and the key
Set up InduQT for this sensor:
1. Create a site called "Plant 1" (skip if it already exists).
2. Create a machine named "compressor-1" with externalId "compressor-1" and family "compressors".
3. Create an ingest key for the site called "Raspberry Pi bridge" and show it to me once.
After I start the bridge, I'll ask you to check the latest values. The signals will be em_0.a_current, em_0.b_current, em_0.c_current, em_0.total_act_power and the energy totals.
2 · Start the bridge
python3 -m venv ~/induqt && source ~/induqt/bin/activate
pip install "paho-mqtt>=2.0" requests

export MQTT_HOST=localhost MQTT_USER=viewer MQTT_PASSWORD='viewer-password'
export MQTT_TOPICS='factory/#'
export INDUQT_API_KEY='iq_live_…'   # the site key Claude or the app created for you
python3 induqt_bridge.py
induqt_bridge.py (the whole bridge, about 120 lines)
induqt_bridge.py (the whole bridge, about 120 lines)
#!/usr/bin/env python3
"""
induqt_bridge.py — forward MQTT sensor messages to InduQT.

Subscribes to your local broker, turns each message into numbers, and posts them
to the InduQT ingest API in batches. Works on every InduQT plan, including the free one,
and your broker never has to be reachable from the internet.

  factory/<machine>/<anything>        ->  machine = <machine> (matched to the machine's externalId)
  factory/pump-3/vibration {"vib_rms_x_g": 0.12}  ->  signal vib_rms_x_g
  factory/comp-1/status/em:0 {"a_current": 12.4}  ->  signal em_0.a_current
      (two or more levels after the machine: the last level names the group, so
       two probes or two channels never overwrite each other)
  {"temperature": {"tC": 41.2}}       ->  signal temperature.tC
  factory/furnace-1/temp  21.7        ->  signal temp (a bare number takes the last level)

pip install "paho-mqtt>=2.0" requests
"""
import json
import os
import re
import threading
import time

import paho.mqtt.client as mqtt
import requests

MQTT_HOST = os.environ.get("MQTT_HOST", "localhost")
MQTT_PORT = int(os.environ.get("MQTT_PORT", "1883"))
MQTT_USER = os.environ.get("MQTT_USER")
MQTT_PASSWORD = os.environ.get("MQTT_PASSWORD")
TOPICS = [t.strip() for t in os.environ.get("MQTT_TOPICS", "factory/#").split(",") if t.strip()]
MACHINE_LEVEL = int(os.environ.get("MACHINE_LEVEL", "1"))  # factory/<machine>/... is level 1
SEND_EVERY = float(os.environ.get("SEND_EVERY", "10"))     # seconds between posts
IGNORE = set(os.environ.get("IGNORE_FIELDS", "id,ts,time,timestamp,unixtime,Time").split(","))
INGEST_URL = os.environ.get("INDUQT_URL", "https://api.induqt.ai/v1/ingest")
API_KEY = os.environ["INDUQT_API_KEY"]  # a site key from InduQT (Settings -> API keys)

BAD = re.compile(r"[^A-Za-z0-9_.\-]")
buffer, lock = [], threading.Lock()


def key(name):
    return BAD.sub("_", name)[:64]


def numbers(prefix, value, out):
    """Collect every number in a JSON value; true/false become 1/0."""
    if isinstance(value, bool):
        out[key(prefix)] = 1 if value else 0
    elif isinstance(value, (int, float)):
        out[key(prefix)] = value
    elif isinstance(value, str):
        text = value.strip().lower()
        if text in ("on", "true", "online"):
            out[key(prefix)] = 1
        elif text in ("off", "false", "offline"):
            out[key(prefix)] = 0
        else:
            try:
                out[key(prefix)] = float(text)
            except ValueError:
                pass
    elif isinstance(value, dict):
        for k, v in value.items():
            if k not in IGNORE:
                numbers(f"{prefix}.{k}" if prefix else k, v, out)


def on_connect(client, userdata, flags, reason_code, properties):
    print(f"MQTT connected ({reason_code}); subscribing to {TOPICS}")
    for t in TOPICS:
        client.subscribe(t, qos=1)


def on_message(client, userdata, msg):
    levels = msg.topic.split("/")
    if len(levels) <= MACHINE_LEVEL:
        return
    machine = levels[MACHINE_LEVEL]
    try:
        payload = json.loads(msg.payload.decode("utf-8", "replace"))
    except ValueError:
        payload = msg.payload.decode("utf-8", "replace")
    values = {}
    group = levels[-1] if len(levels) - MACHINE_LEVEL > 2 else ""
    if isinstance(payload, dict):
        numbers(group, payload, values)
    else:
        numbers(levels[-1], payload, values)
    if values:
        with lock:
            buffer.append({"machine": machine, "ts": int(time.time() * 1000), "values": values})


def sender():
    while True:
        time.sleep(SEND_EVERY)
        with lock:
            batch = buffer[:500]
        if not batch:
            continue
        try:
            r = requests.post(
                INGEST_URL,
                headers={"Authorization": f"Bearer {API_KEY}"},
                json={"readings": batch, "autoCreate": True},
                timeout=15,
            )
            if r.status_code < 300:
                with lock:
                    del buffer[: len(batch)]
                print(f"sent {len(batch)} readings: {r.json()}")
            elif r.status_code in (400, 401, 403):
                print(f"InduQT refused the batch ({r.status_code}): {r.text}")
                with lock:
                    del buffer[: len(batch)]  # fix the key or mapping; retrying will not help
            else:
                print(f"InduQT returned {r.status_code}; will retry")
        except requests.RequestException as e:
            print(f"network error, will retry: {e}")
        with lock:
            del buffer[:-5000]  # keep at most ~5,000 readings while offline


client = mqtt.Client(mqtt.CallbackAPIVersion.VERSION2, client_id="induqt-bridge")
if MQTT_USER:
    client.username_pw_set(MQTT_USER, MQTT_PASSWORD)
client.on_connect = on_connect
client.on_message = on_message
threading.Thread(target=sender, daemon=True).start()
client.connect(MQTT_HOST, MQTT_PORT, keepalive=60)
client.loop_forever()
Keep it running after a reboot (systemd)
Keep it running after a reboot (systemd)
# /etc/systemd/system/induqt-bridge.service  — keeps the bridge running after a reboot
[Unit]
Description=MQTT to InduQT bridge
After=network-online.target mosquitto.service

[Service]
User=pi
EnvironmentFile=/home/pi/induqt-bridge.env
ExecStart=/home/pi/induqt/bin/python3 /home/pi/induqt_bridge.py
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

# then:
# sudo systemctl daemon-reload && sudo systemctl enable --now induqt-bridge
# journalctl -u induqt-bridge -f

The bridge treats the topic level after factory/ as the machine (it matches the machine's externalId and creates it if needed), turns every number in the message into a signal (nested fields become names like temperature.tC; when the topic has a group after the machine, such as Shelly's status/em:0, the group goes in front, as em_0.a_current, so two channels never overwrite each other; true/false and on/off become 1/0) and sends a batch every 10 seconds. It was run end to end on 2026-10-31 against Mosquitto 2 with Shelly Gen2, Theengs and plain-number messages. Prefer Node-RED? The same bridge as a Node-RED flow is on the gateway guide.

Prompt
3 · Check it arrived and set what normal looks like
Show me the latest values InduQT has for compressor-1.
Then set these normal bands so InduQT knows what "normal" looks like: em_0.a_current between the unloaded and the full-load current from the nameplate.

On the Plant Monitor plan InduQT can also subscribe to a broker directly. Point your sensors (or a Mosquitto bridge) at a cloud broker with TLS, then ask Claude to "create an MQTT connector for this broker, machine from topic level 1", or add it in the app under the site → Connect data → MQTT broker.

What InduQT does with it

  • Separates running loaded, running unloaded and stopped, and learns how long each lasts at this site.
  • Flags phase imbalance and a rising unloaded current, two early signs of electrical and mechanical trouble.
  • Compares this week’s loaded hours with the same demand last month, so leaks show up as a trend, not a surprise bill.
  • When something changes, an investigation names the signal that moved first and the likely causes, from the compressor’s manual if you add it.
Prompt
Ask Claude after a week
In InduQT, analyze compressor-1 over the last 7 days. How many hours was it loaded vs unloaded each day, how many starts per hour at peak, and is anything drifting?

Frequently asked questions

Can I tell loaded from unloaded with only a current clamp?

Yes, for most fixed-speed screw compressors. Loaded current is close to full-load amps and unloaded drops to roughly a quarter to two fifths of it, so the two states are clearly separate. Variable-speed compressors show a continuous range instead, which InduQT treats as load.

Do I need all three clamps?

One clamp gives load and run hours. Three show phase imbalance and give accurate power, which is worth the small extra cost.

What about pressure and temperature?

Add a Shelly Plus Uni with a 4–20 mA pressure transmitter (see the hydraulic pressure guide) or DS18B20 probes on the discharge line (see the furnace cooling guide). Both publish over MQTT the same way.

Related