Platform · AI Assistant and industrial AI

Ask your plant. The AI Assistant proposes; you approve.

Proxus AI Assistant is built into the hub. Ask from any screen in plain words and it answers from the live plant, with your permissions. It prepares changes, and nothing is applied until you approve. Your own assistants and agents reach the same governed model through an MCP interface, under the same rules.

Book a demo Plan your assistant

Screen recording: the AI Assistant opened with Ctrl+K, answering what needs attention in the plant, then proposing an alarm rule that waits for approval.

AI Assistant: Built into the hub: Monitor > AI Assistant, or Ctrl+K on any screen
Changes: Proposed on a card and made only after you approve, in your name in the audit trail
AI service: Claude, OpenAI, Azure OpenAI, Claude on Amazon Bedrock or Google Vertex AI, or a local Ollama model
MCP: A Model Context Protocol server in your deployment, for your own assistants and agents

From a question to a change

It reads with your permissions and changes nothing on its own.

Ask in plain words; the AI Assistant reads the live plant as you, and a change comes back as a card you approve.

Question “Warn when mixer 2 stays above 180 A for 10 minutes”
Reads Devices, alarms, history Only what your role may see
Proposal A new alarm rule, tried on stored data Waiting for your approval
Applied After you approve In your name in the audit trail

In the product

Proposals, permissions, your own agents.

It proposes, you approve

A change comes back as a card with every field it would set. Approve and apply, or reject; until then nothing changes.

A new alarm rule is tried on stored data first
An unanswered proposal expires after 24 hours

A role’s AI Assistant settings: its access and daily token limit.

Enlarge

Only what the role allows

Each role sets what the AI Assistant may do for its people: nothing, answer questions, propose changes, or also run actions, with a daily token limit.

A delete needs Approve and delete, then a second confirmation
Tokens and estimated cost per AI service and per user

AI Assistant usage: answers, tokens and estimated cost per service and per user.

Enlarge

What it costs, per service and per user

Administrators see the month’s answers, tokens and estimated cost, per AI service with its monthly budget and per user.

Prices per million tokens set on each AI service
Usage records stay when a conversation is deleted

A real MCP session against the demo hub

POST /mcp  initialize
← Proxus hub · tools, prompts, resources

POST /mcp  tools/list
← 55 tools: list_devices, telemetry_query, tag_history,
  open_alarms, propose_create, …

POST /mcp  tools/call auth_login
  { "userName": "Admin", "password": "…" }
← success · Admin · session for 25 minutes

POST /mcp  tools/call device_status  { "deviceOids": [66] }
← MFB_LYS_MIXER_3 · Online · Connected · gateway online

Your own agents, over MCP

Any MCP client (Claude Desktop, MCP Inspector, your own agent) opens a session at /mcp on the hub, signs in as a Proxus user and calls the same tools, within that user’s permissions.

Writes stay off until an administrator allows them
An agent’s change arrives as a card to approve

Real Proxus screens; the machine data shown is simulated.

Want to ask it about your own plant? Plan your assistant

Technical details Problems it solves, every capability and the full protocol table

What it solves

What teams run into without it.

AI on stale exports

An assistant answering from last quarter’s spreadsheet is confidently wrong about this morning’s line.
No boundary on what it can see

Handing a model a database credential is not an access policy.
No record of AI access

Without an audit trail, nobody can say who used which tool or what data left the plant.

What you get

Built in, not bolted on.

AI Assistant on every screen

Ask in plain words why a device is silent, what changed or which alarms are open. Each answer names the sources it read: devices, alarms, tag history, the audit trail, the docs.
Your permissions, nothing more

The AI Assistant reads with the asking user’s role and plant-area limits. Passwords, keys, tokens and licence details are never read.
Changes you approve

Any record you may edit comes back as a proposal card, one at a time, several at once or as one plan. A new alarm rule is first tried on stored data, so you see how often it would have fired. An approved change is made with your permissions and logged under your name.
Your choice of model

Claude, OpenAI, Azure OpenAI, Claude on Amazon Bedrock or Google Vertex AI, or a local Ollama model when no data may leave the site. Monthly budgets per service, daily token limits per role.
Governed MCP server

Your own assistants and agents query the same model as tools. MCP can be switched off, restricted to administrators, limited to read methods and rate-limited; every request is logged.
Insights without asking

Noisy alarms, alarms nobody owns, unstable gateway links, frozen or unusual values and expiring certificates, found by fixed background checks. No AI service is used until you ask the AI Assistant about one.

What this enables

What the capabilities add up to, in practice.

Ask about the plant in plain words and get answers from live data, within your permissions
Let the AI Assistant prepare rules and changes that are applied only after someone approves them
Ground AI answers in current production data instead of stale exports
Give an assistant a permissioned scope rather than a database credential
Switch AI access off, restrict it to administrators or read methods, and rate-limit it
Review every request by user, tool and result, like any other integration

Questions

What teams ask before the first call.

Ask something else

What is the Proxus AI Assistant?

An assistant built into the hub. Ask in plain words from any screen (Ctrl+K) or in Monitor > AI Assistant. It answers from the plant data you may read, names the sources it used, and prepares changes as proposal cards that are applied only after you approve them, in your name in the audit trail.

Which AI model does the AI Assistant use?

The one you choose: Claude, OpenAI, Azure OpenAI, Claude on Amazon Bedrock or Google Vertex AI, or a local Ollama model, with which nothing leaves your site. Administrators set a monthly budget per service and a daily token limit per role.

What is MCP and why does it matter here?

Model Context Protocol is an open interface that lets AI assistants call defined tools instead of scraping data. It matters because it turns "give the model database access" into "give the model a permissioned, audited set of operations".

Can AI change anything in the plant?

Only what the user is already allowed to do. The AI Assistant changes nothing on its own: each change is a proposal card that is applied only after you approve it, with your permissions. Changes that create or edit records can be undone; deletes and alarm decisions cannot, and the card says so before you approve. MCP offers read tools plus one write tool for create and update operations, limited to the signed-in user’s role and namespace-path permissions; deletes are blocked by default, and MCP can be limited to reading.

Does our production data leave the network?

That depends on the AI service you choose for the AI Assistant. With a cloud service, each question sends the question, the conversation so far and the plant data the assistant read to answer it (device names and states, latest tag values, alarms), and only what the asking user may read. With a local Ollama model, nothing leaves the site. With Amazon Bedrock, Google Vertex AI or Azure OpenAI, the data stays within your own cloud account. MCP runs inside your deployment; what an AI client you connect to it sends onward is your policy.

5 more questions

How do we limit what an assistant can see?

The same way you limit a person: roles and namespace-path permissions. MCP can also be switched off, restricted to administrators, limited to read methods and rate-limited.

Is AI usage auditable?

Yes. A change approved from the AI Assistant is in the audit trail under your name, with the assistant and the conversation as its source; conversations are kept in the hub, and AI Assistant Usage shows answers, tokens and estimated cost per service and per user. Every MCP request is logged with user, tool and result, like any other integration.

Does the AI Assistant work without internet access?

Yes, with a local Ollama model. The documentation the assistant searches ships with the hub, so it works offline too.

What does the AI Assistant cost to run?

The assistant uses your own key or cloud identity, so your AI provider bills you directly. Administrators set a monthly budget per service and a daily token limit per role.

Can the AI Assistant be wrong?

Yes. Each answer names the sources it read, so an important one can be checked on the screen it names, and nothing changes without your approval.

Continue

Two numbers set your price: the gateways you run and the data rate they carry. See pricing

Ask your plant. The AI Assistant proposes; you approve.

It reads with your permissions and changes nothing on its own.

Proposals, permissions, your own agents.

It proposes, you approve

Only what the role allows

What it costs, per service and per user

Your own agents, over MCP

What teams run into without it.

AI on stale exports

No boundary on what it can see

No record of AI access

Built in, not bolted on.

AI Assistant on every screen

Your permissions, nothing more

Changes you approve

Your choice of model

Governed MCP server

Insights without asking

What the capabilities add up to, in practice.

Running in production at these plants.

What teams ask before the first call.

Which model, and where may the data go?