AI agents can now help build Power Apps apps

You can now have an AI agent help build your app in Power Apps. The feature is a plugin for so-called canvas apps, which, according to an article dated September 4, 2026, has been declared generally available for production use. A canvas app is an app that you put together screen by screen in Microsoft's tool.

The idea is that the user and the agent work in the same app. The user can ask the agent to create screens, connect data or update formulas, and then review and correct the result. The agent can also write Power Fx formulas (Power Apps' formula language), configure controls and validate changes.

Users choose which coding agent to use, and they have to keep an eye on their own token usage, the unit AI services typically use to measure consumption. You can switch between working with the agent's help and editing directly.

The responsibility still lies with the user

According to the article, the solution is meant as an aid that gives makers more leverage, not as a replacement for them. Makers remain responsible for governance, security, testing, accessibility and deployment of the app.

How it works technically

According to the documentation on GitHub, the plugin uses a so-called Canvas Authoring MCP server. It lets agents generate, validate and compile the app's YAML files (.pa.yaml) together with a running session in Power Apps Studio. The server requires the .NET 10 SDK to be installed locally.

The Power Apps Studio tab in the browser must stay open throughout the session. If it is closed, the session stops, and compiling and syncing no longer work. The agent must also sync before it edits the local files, so that outdated changes are not introduced.

When a new app is created, five specialized agents are involved: functionality, screens, layout, discovery and builder. When editing, a planner and a screen builder are used.

Test: A good starting point, but not all the way

In a hands-on test from HubSite 365, the author used SharePoint and Dataverse data in a marketing app. The agent contributed to the initial build of the app, but manual intervention was needed for OData fields and loops and for a problem with saving.

According to the test, checking and manual correction must still be part of the process. The test is based on one person's experience. The feature must be turned on in the admin center. The sources do not say which licenses, regions or prices apply.