Your AI agents — LLM-based operator assistants, optimisation agents, alarm triage — need current plant data and a controlled way to act. ConnectMQ gives them both: they subscribe to the unified namespace, reason over live values and publish proposals only on topics they are explicitly allowed to use.
No card required
PUBLISH agents/quality-agent/proposals
{"id":"p-2291","asset":"acme/dallas/molding/line-2/press-04","action":"setpoint","tag":"barrel_temp_z3","from":238,"to":232,"reason":"short shots rising","approval":"required"}QoS 1 · TLS 8883 · PUBACK
Most AI pilots in plants run on exported CSVs or a nightly copy of the historian, so their advice is hours old. The opposite shortcut — giving an agent direct access to SCADA or PLCs — is a risk no OT team will sign off. What is missing is a real-time data layer with clear limits on what an agent may read and what it may write.
PLC gateways, MES and quality systems publish into an ISA-95 topic tree on ConnectMQ: states, process values, alarms and orders, with current values retained.
Each agent connects with its own credentials and subscribes to the area it covers — for example one line’s alarms, process values and the current production order.
The agent publishes a recommendation, such as a setpoint change or a maintenance request, to its proposals topic. An operator approves it in your HMI or app, which publishes the command.
Only the approval path may write to command topics. Proposals, decisions and outcomes are published to state topics and stored by your historian or log, so every action is traceable.
Per-topic ACLs let an agent read a whole area but publish only to its own proposals topic. It cannot write to command topics unless you allow it.
Retained messages give an agent the current state of every asset the moment it subscribes, without querying each system separately.
Agents use any MQTT library — Python, Node.js, Go — over TLS or secure WebSocket, just like gateways and dashboards.
The live topic monitor shows exactly what each agent publishes, and Last Will marks an agent offline when its process stops.
acme/dallas/molding/line-2/+/state (retained)
acme/dallas/molding/line-2/alarms
agents/quality-agent/proposals
agents/quality-agent/status (retained)Yes. An agent is just another MQTT client: your agent runtime uses a standard library such as paho-mqtt or MQTT.js to subscribe and publish, usually exposed to the model as tools like “read line status” or “propose setpoint”. ConnectMQ provides the broker, credentials and ACLs; the agent and the model are yours.
Give the agent credentials whose ACL allows publishing only to a proposals topic, never to command topics. A human-in-the-loop step in your HMI or app — or a validation service you trust — publishes the actual command. PLC interlocks and safety systems stay exactly where they are.
Not strictly, but it helps a lot. A UNS gives agents one consistent, real-time view of the plant through one interface, instead of separate integrations with SCADA, MES and the historian. With MQTT, the same namespace serves agents, dashboards and analytics.
Start on the Free plan — no card required — connect your first devices today and upgrade when you go live.
