Building Power Apps With Claude Code or GitHub Copilot: The Canvas Authoring Agent Plugin

In Short: Build Power Apps With Claude Code or GitHub Copilot

You can now build Power Apps by describing them to a coding agent. Microsoft's canvas authoring agent plugin, announced as generally available on 17 September 2026, lets GitHub Copilot CLI, Claude Code and other coding agents that support MCP create and edit canvas apps. The agent writes the app as code, checks it against Microsoft's canvas authoring service and syncs it into Power Apps Studio while you watch.

At the same time Microsoft announced a model-driven app builder skill, also generally available, which plans and builds a whole model-driven app in Dataverse from a description of the business process.

This is a real change in how Power Apps gets built. It is also a change that needs the same guardrails as any other way of putting apps in front of users. Below is how the canvas plugin works, how to set it up, and what we would put in place before handing it to makers.

How the Canvas Authoring Agent Works

The plugin does not replace Power Apps Studio. It works alongside it, through a live coauthoring session:

• You create and save a canvas app in Power Apps, in the right environment. It can be blank, but it must exist, and coauthoring must be switched on (Settings, then Updates, then Coauthoring).
• You connect your coding agent to that Studio session through the canvas app authoring MCP server.
• You describe what you want, for example "a canvas app for tracking expense reports with an approval workflow". The agent can ask follow-up questions.
• The agent finds the available controls, connectors and data sources, then writes .pa.yaml files that define the screens, controls and Power Fx formulas.
• It validates those files against the MCP server, fixes any errors, and the changes sync into Studio.

Editing works the same way. The agent pulls the current app from the coauthoring session into local .pa.yaml files, makes the changes you ask for, validates them and syncs them back.

The plugin has three skills: /canvas-app to build or edit, /configure-canvas-mcp to connect to the Studio session, and /add-data-source to guide you through adding a connector or data source and checking it is available.

One detail matters for planning: the connection depends on the open Studio tab. If you close it or the session expires, you reopen the app and connect again. This is a working session with a person present, not a background build server.

Setting It Up

The requirements are short:

• .NET SDK 10.0 or later. The MCP server runs as a .NET package, so the full SDK is needed, not just the runtime. Check with dotnet --list-sdks.
• A coding agent that supports MCP servers and agent plugins, such as GitHub Copilot or Claude Code. You need whatever subscription that tool requires.
• A Power Platform environment where you can create or edit canvas apps.

In a coding agent with plugin marketplace commands, you add the microsoft/power-platform-skills marketplace and install the canvas-apps plugin. In Visual Studio Code, search the Extensions view for the Canvas Apps agent plugin published by Microsoft. Then copy the Power Apps Studio URL from your browser (the edit session, not the app player), run /configure-canvas-mcp, paste the URL and sign in. A quick check is to ask the agent to list the available canvas app controls.

Two set-up problems will be familiar to anyone who works in a locked-down tenant. Organisational policies can block agent plugins, third-party MCP servers or local MCP processes. And the MCP server downloads its package from NuGet, so if your organisation only allows an internal package feed, your administrator needs to make the package available there.

The Wider Plugin Family

The canvas plugin is one part of Microsoft's power-platform-skills repository on GitHub, which supports GitHub Copilot CLI and Claude Code and is MIT licensed. Alongside canvas apps it includes plugins for model-driven apps, code apps, Power Pages, Power Automate cloud flows and more.

The model-driven app builder is the one to watch for system-shaped apps. You describe the business scenario, the skill proposes an app spec built around personas and the jobs they need to do, then shows a dry-run build plan. Microsoft is explicit that no app artifacts are created until you approve that plan. After approval it creates tables, columns, relationships, forms, views, charts, business rules, business process flows, generative pages, a sitemap and security roles in Dataverse, and can check the deployed app against the approved spec. It needs Node.js, the Power Platform CLI and the Azure CLI, signed in with the same identity.

If you are unsure which kind of app a problem needs, our guide to what Power Apps is and the two kinds of app covers the canvas versus model-driven choice.

What Changes for Makers and IT

The obvious gain is speed. A first working version of an app can now come from a conversation rather than an afternoon of placing controls and writing formulas. That helps with the long tail of requests IT never reaches, the problem we described in clearing the IT backlog with low-code.

The less obvious change is that canvas apps now exist as readable files on the developer's machine. That is the same shift Power BI went through with PBIP and PBIR: once the app is text, it can be reviewed, compared and kept in source control. Teams that already use coding agents for data work, as in our post on agent skills for Fabric and the GitHub Copilot CLI, will find the workflow familiar.

What does not change is the data decision underneath. A coding agent will happily build a polished app on a SharePoint list that should have been a Dataverse table. Make that choice deliberately first; our SharePoint lists versus Dataverse guide sets out the trade-offs.

The Guardrails We Would Set

Microsoft's own security guidance for the plugin is sensible, and we would turn it into rules:

• Build in development environments. Microsoft advises validating generated changes in a development or non-production app before applying them to production.
• Least privilege. Connect with an account that has only the permissions needed to edit the intended app.
• Read before you approve. Check the environment, app URL and proposed action before approving an operation, and review the generated .pa.yaml files against your standards.
• No secrets in prompts. Never paste passwords, tokens or connection strings into the conversation.
• Data policies still apply. Connectors are added through Studio and your data loss prevention policies can restrict them.
• Watch for prompt injection when the agent works with external data, and review unexpected actions.

Microsoft also notes that the generated code is a best-effort attempt and that responsibility for accessibility, security and compliance stays with you. In practice, that means a person who understands Power Apps still reviews every app before users see it.

Where Solv Systems Comes In

We build Power Apps and the governance around them, and AI-assisted authoring is now part of how we work. The value we add is less in typing prompts and more in the parts that decide whether an app survives: the data model, environment strategy, data policies, review steps and ownership. If you want your makers using Claude Code or GitHub Copilot without creating a new generation of unmanaged apps, our Power Platform services are the place to start.

Sources and Further Reading

• What's new in Power Platform: September 2026 feature update
• Create and edit canvas apps with AI code generation tools
• Build and edit model-driven apps with AI code generation tools
• Power Platform Skills on GitHub