Starter prompt
Scaffold a page
A new page and a starter widget tree from a one-line brief, with bindings checked against the widget schemas.
“A line overview with a status header and four KPI cells.”
AI-assisted engineering
Let an AI assistant do the repetitive part.
Forty alarms from a spreadsheet, twenty near-identical screens, every string on a page into the translation table. NEXT HMI has a built-in MCP server, so an assistant such as Claude can do that work in your project. Every change comes back as a diff, and anything destructive waits for your confirm.
How it works
An assistant that edits the same files you do.
The MCP server (Model Context Protocol) exposes 30+ read and write tools over the same pages, datasources, alarms and translations the editor writes. There is no separate export format and no scripting layer on the side. Every write returns a diff of what changed, and a destructive change comes back as a dry run first. Because the project is plain files, forty near-identical screens are reviewable the way a colleague's pull request is.
The assistant talks straight to your own server. Your project files never pass through a NEXT HMI service, and there is nothing to sign up for.
What it's good at
Four jobs, ready to run.
These ship with the server as starter prompts. Once your assistant is connected, they appear in its prompt menu, so you don't have to describe the workflow yourself.
Starter prompt
Dashboard from a datasource
Walks a datasource's variable tree and lays out the values that matter on a new page, up to the number of widgets you set.
“A dashboard for the LinePLC datasource, at most twelve widgets.”
Starter prompt
Seed alarms
Picks the boolean variables whose names point to a fault and writes one alarm definition for each into a group you already have.
“Seed alarms from LinePLC into the group Line faults.”
Starter prompt
Localize a page
Moves every static string on a page into a dictionary and rewrites the page to read them from it.
“Localize the line overview into the Default dictionary.”
What it can and can't touch
The limits are in the server, not in the prompt.
An assistant can't talk its way past these. They hold whatever it is asked to do.
| Area | An assistant may |
|---|---|
| Pages, widget trees, widget properties | read + write |
| Global events, variable triggers included | read + write |
| Variables on a datasource | read + write |
| Alarm definitions | read + write |
| Translations: languages, keys, cells | read + write |
| Assets: icons and images | read + write |
| Datasource connections and browsed PLC trees | read only |
| Components and the widget-schema manifest | read only |
| Users: names, groups, enabled state | read only, never credentials |
| Starting, stopping, creating or deleting projects | never. That stays an operator action in the dashboard. |
The full tool catalogue, argument shapes and limits are in the MCP guide.
How you stay in control
Off until you turn it on. Scoped when you do.
Every gate below is in the server. None of them depends on the assistant behaving.
- Opt-in per project. A checkbox on the Manager dashboard. Off by default: an assistant can read the project, but every write is refused.
- One token, one project. Each token is for a single project and is either read or write. Revoke one by id, or all of them by changing the device-admin password.
- A diff before anything destructive. The first call returns the change. Only a second, confirming call applies it.
- Operator screens follow along. A running project re-reads the changed files and pushes them to every connected browser.
- Your editor asks first. An open editor shows a banner and lets you reload or dismiss. It never swaps the canvas while you are working.
Connect
Three steps, any MCP client.
- Tick MCP enabled on the project's row in the Manager dashboard.
- Pair a token with the device-admin password. It is shown once; only its hash is stored.
- Add the server to your assistant. Claude Code is shown here. Clients configured by file take the same URL and header.
On Windows PowerShell, with HTTPS, or from another machine: the MCP guide has each variant.
# 2 · pair a write token for project line-a curl -X POST http://localhost:8000/api/manager/mcp/pair \ -H 'Content-Type: application/json' \ -d '{"password":"…","projectId":"line-a", "access":"write","name":"Claude"}' # 3 · add the server to Claude Code claude mcp add --transport http nexthmi \ http://localhost:8000/mcp/ \ --header "Authorization: Bearer <token>"