The AI assistant

What is it?

One assistant, always in the bottom-right corner of Studio (the round sparkle button). It has two jobs, on two tabs:

Tab	For	Works on
Edit this	changing what you have open	the page or form open in its designer
Create	making something new	an entity, dashboard, form, page, integration flow, workflow, theme, or a whole plugin

A small dot on the button means something is open that AI can edit. The panel opens on Edit this when a page or form is open, otherwise on Create. Your conversation is kept when you switch tabs.

When to use it

Use it to get a first draft fast, to make a repetitive change, or when you know what you want but not which screen makes it. Do not use it to skip review: everything it makes is a draft you check.

Edit this: change what is open

1. Open a page or form in its designer.
2. Click the sparkle button. The header says "Editing this page: Loans".
3. Type what you want changed, in plain words.
4. AI answers with a list of proposed changes. Each one is a real designer command (add a field, move a block, set a property).
5. Tick the ones you want and press Apply. Untick any you do not.
6. Look at the canvas. Ctrl+Z undoes an applied change like any other edit.

Examples that work well

You type	You get
Add a GST Number field below Customer Name	a text field, positioned under Customer Name
Make Reason required when Leave type is Unpaid	a form rule (output required)
Add a Status filter above the table	a select block bound to a filter variable
Change the page title to Team roster	one property change
Add two KPI cards above the table: Pending and Approved	two card blocks in a new row

Be specific about names (the field or block you mean) and position (above, below, in the second column).

Create: make something new

1. Click the sparkle button and choose Create.
2. Describe it. Example: create an entity for tracking company vehicles.
3. AI routes your request to the right Studio capability and tells you exactly what it made.
4. Use the Open in Designer chip in the reply to jump straight to it and review.
5. Publish only after you have checked it.

Examples

You type	What is made
create an entity for tracking company vehicles with plate number, model, driver and service due date	an entity with those fields, ready to refine in the Entity Designer
create a leave approval workflow: manager then HR, escalate to HR after 2 days	a draft workflow with two stages and an SLA
create a dashboard with pending, approved and rejected leave counts	a draft dashboard
create a page listing equipment with a search box and a Lend button	a draft page
create a calm blue theme for the facilities desk	a draft theme

There is also a microphone button beside the box: speak instead of typing.

Build a whole plugin with words

In the Create Plugin wizard, step 1 has Describe your plugin. Type:

Equipment loans for the facilities team: track each loan with item, borrower, due date and status,
manager approval for anything over 500, and a report of overdue loans

Choose Draft with AI. The wizard closes and the assistant opens with your request ready in its input; nothing is created until you send it. Then check the result in the Explorer, run Validate, and continue with Check, build and publish.

A full example: Leave Management in six prompts

Send these one at a time on the Create tab, checking each result:

1. create an entity lm_leave_request with employee, leave type, from date, to date, days, reason and status draft/pending approval/approved/rejected
2. create an entity lm_leave_type with code, name and days per year
3. create a workflow leave-approval with a manager stage and an HR stage
4. create a page listing leave requests with a status filter
5. Open that page, switch to Edit this: add a New request button above the table that opens the leave request form
6. Open the form, Edit this: make Reason required when Leave type is Unpaid

Then Validate and publish. The finished result should look like the capstone.

Can AI design a whole plugin, end to end?

Partly, and it stops before publishing. The assistant drives the same builder Studio uses. For a plugin made of configuration (data, screens, menus, translations, and optionally data services, providers, workflows, roles, permissions, rules and data views) it can do the whole first draft in the order a person would:

scaffold the plugin -> add entities -> add pages -> add the menu -> add translations -> (optional pieces) -> validate -> build

and it reports what it made and that it is built and ready for your review. It never publishes or installs. You then review in the Explorer, run Validate, add what it could not, and press Publish yourself.

Step	AI can do it	You still do it
Create plugin, entities, pages, menu, translations	yes	review and refine in the designers
Data services, providers, workflows, roles, permissions, rules, data views	yes, when your request implies them	check names and approvers
Forms, dashboards, themes, reports, print templates	yes, one by one on the Create tab	connect them to data
Validate and build	yes	read the result
Publish, install, roll back	no	always you
Java or Node backend code	not yet (see below)	write it, or ask an external coding assistant

Quality depends on the model your tenant has configured. A small local model gives rough drafts; a stronger model gives better ones. Always review.

Can AI write backend and frontend code?

The platform's backend languages are Java, Python, Node.js and Go, and the frontend can be React, Vue, Angular, Svelte or plain JavaScript (any framework that can produce a Web Component; see Build a full-code frontend for a plugin). Here is exactly what AI can do with each today.

	Studio assistant (chat in Studio)	An AI coding assistant through MCP or the CLI
Metadata (data, screens, workflows, rules...)	yes, see above	yes (artifact-*, rule-*, entity-* operations)
Java, Python, Node.js, Go, React source files	not yet: it has no tool that writes source files	yes: source-files, source-read, source-write
Compile and test	not yet	Java: java-compile, java-test, java-package; Node: node-run; Python: run pytest yourself; Go: docker build (it vets and tests) or go test
Starter projects	not yet	erp_service_plugin_scaffold (Python FastAPI, Node.js Express or Go net/http), erp_plugin_scaffold (extension points), Add Java backend
Publish and install	no	no: always you

So today the way to have AI write code is an AI coding assistant (Claude Code, an IDE agent) connected to the MCP server: it scaffolds the service, writes the files into the plugin, compiles and tests, and you review and publish in Studio.

Planned: the same file tools for the Studio chat assistant, so "add an endpoint that calculates leave accrual in Python" becomes one request inside Studio.

Keep logic in configuration (rules, workflows) where the assistant is strongest, and use code only where needed (Code extensions).

Tips

• One request, one purpose. Several small prompts beat one long one.
• Name things the way they are named in Studio (the exact field, block or page title).
• If the answer is wrong, say what to change ("use a select, not a text field"); it keeps the conversation's context.

Security and control

• AI acts as you: it can only do what your account may do. Treat everything it makes as a draft, and publish only after your own review.
• Changes on the Edit tab arrive as commands you accept one by one; creations arrive as drafts.
• Every AI action is recorded with your other work in the audit trail.
• The assistant uses the model and API key your tenant sets in AI provider settings (add your own key on the API keys tab; see AI Designer). Cost is billed to your own provider account.

What it does not do (yet)

It edits only pages and forms in place (other artifacts are created by the Create tab and then edited by hand); it does not manage access policies; and results depend on the configured model, so always review.

Common problems

You see	Fix
The Edit this tab says to open a designer	open a page or form first
The Create tab answers with an error or nothing happens	the tenant has no Studio assistant set up yet (the studio-copilot-agent and its tools), or no AI provider: ask your administrator, see AI provider settings. The Edit this tab uses separate endpoints and may still work
The result is not what you meant	reply with the correction, or edit in the designer
The change is not visible	you have not pressed Apply, or it is a draft not yet published

For agents and scripts

The same operations are available outside the browser: see Command line and MCP.