Status: Public Preview — behaviour may change.

What it does

The model app builder skill lets you build and edit model-driven apps from a plain-language description inside an AI code generation tool such as GitHub Copilot CLI or Claude Code, rather than assembling each table, form, view and page by hand. You describe the process, the skill turns it into an app spec, you approve a dry-run build plan, and it writes the app plus its supporting Dataverse artifacts into your environment.

Key facts

• Runs outside make.powerapps.com — it's a plugin for a command-line AI code tool, driven through a /app-builder command, not a designer surface.
• Prerequisites are real dev-machine tooling: Node.js (current LTS), Power Platform CLI 2.7.0+, Azure CLI, and a supported AI code tool. You need an authenticated PAC CLI profile against the target environment and a matching az login session under the same identity.
• Install is via the Power Platform Skills marketplace: add microsoft/power-platform-skills, then /plugin install model-apps@power-platform-skills. A one-line installer script sets up all the plugins and turns on auto-update.
• It produces standard artifacts — tables, columns, relationships, forms, views, charts, sample data, a sitemap with per-table icons, JavaScript validation rules (as web resources), generative pages, and security roles based on the planned personas. Nothing proprietary; you can keep editing everything in Power Apps afterwards.
• Multistep and gated: describe scenario, review app spec, review dry-run build plan, build, optional verification against the approved spec, iterate. No artifacts are written until you approve the plan.
• It doesn't yet cover every model-driven artifact or concept — unsupported components are left to the standard designers.
• Generative-page-only work has its own separate skill; use that if you're not building a whole app.

When to use / skip

Reach for it when you're standing up a new app from a reasonably well-understood process and want the scaffolding — schema, forms, views, navigation, a security model — generated in one pass instead of clicked out over a day. It's also handy for bulk edits described in a sentence. Skip it for small, surgical changes to a mature app where the designers are faster and lower-risk, and skip it entirely if you're not comfortable running PAC CLI, Azure CLI and a command-line AI tool — this is a maker feature with a developer's setup. It's preview, so keep it off anything you can't afford to have rebuilt.

Configuration decisions

• Which AI code tool you standardise on (GitHub Copilot CLI, Claude Code, or another supported one) and, for Claude Code, the plugin scope — global, local or user — since that dictates which directory you have to be in.
• The identity behind the PAC CLI profile and az login, which is what the skill builds as — get the environment and privileges right before you start.
• How much detail to put in the initial description: tight requirements when you know exactly what you want, a broader brief when you want the skill to propose a design.
• Whether the optional verification step is part of your build routine.

Gotchas

• PAC CLI and Azure CLI must be authenticated as the same identity — a mismatch is the obvious first thing to trip over.
• Changes you make in the designers between AI-assisted builds can be affected by the next build; review the regenerated app spec and plan before approving so manual customisations survive.
• Auto-update is on by default and Microsoft ships often during preview, so the skill's behaviour and its app-spec schema can shift under you between runs — worth pinning versions if you're mid-project.
• The dry-run plan is the only safety gate. Approving without reading it is how unwanted tables and roles land in an environment.
• It writes to a real environment via an authenticated profile — point it at a dev environment, not production, until you trust the output.

Consultant notes

• Clients hear "AI builds the app" and scope out the maker effort entirely. It generates a solid first pass, but personas, data model and security still need a human who understands the business — budget for review and rework, not zero-touch delivery.
• The prerequisite stack (Node, PAC CLI, Azure CLI, a CLI AI tool, two authenticated sessions) is a genuine barrier for a citizen-maker audience. In practice this lands with pro-makers and consultants, not the business users the "describe your app" pitch implies.
• Preview means no production commitment and a moving schema. Don't build a delivery methodology around it yet; use it to accelerate prototyping and discovery, then harden the output with the standard tooling.
• Because it emits standard Dataverse and Power Apps artifacts, there's no lock-in — but that also means the usual ALM discipline applies. Get generated apps into a solution and source control early, before the next AI edit reshapes them.

---

Preview capability under active development — revisit after the next release wave or when it reaches general availability, as the app-spec schema and command options are expected to change.