Microsoft Power Apps MCP Server Is Now in Public Preview — AI Agents Inside Your Business Apps, With Humans in Control
Microsoft's Power Apps MCP Server brings AI agents into model-driven apps with human-in-the-loop supervision, automated data entry, and a redesigned agent feed.
Microsoft has released the Power Apps MCP Server into public preview, and it represents a meaningful shift in how AI agents can work inside model-driven Power Apps. This is not a chatbot bolted onto a sidebar. It is a structured way for AI agents to take actions inside your business applications, with humans able to review, approve, or take over at any point.
What Is MCP, and Why Does It Matter Here?
MCP stands for Model Context Protocol, an open protocol that gives large language models a standardised way to connect to external tools and data sources. Think of it as a common plug socket for AI agents. Instead of every AI integration requiring its own custom wiring, MCP provides a consistent interface.
The Power Apps MCP Server uses this protocol to let a Microsoft Copilot Studio agent communicate directly with your model-driven Power App. That means agents can now interact with your Dataverse data and your app’s forms in a controlled, auditable way, without any bespoke integration code.
Three Tools, Three Modes of Working
The Power Apps MCP Server gives agents three tools to work with, each suited to a different level of human involvement.
invoke_data_entry is the one most people will reach for first. An agent receives some unstructured input, say an email or a scanned document, analyses it, extracts the relevant field values, and populates a Dataverse form automatically. The completed form then appears in the agent feed as a task waiting for a human to review and approve. Supported file formats include PDF, Word, Excel, and common image formats. The agent does the tedious extraction work; a human confirms it before anything is saved.
request_assistance is for situations where the agent hits a wall and genuinely needs a person to step in. It creates a blocking task in the agent feed and waits until a human completes the required action before the agent continues. This is the appropriate tool when the agent cannot proceed safely on its own.
log_for_review sits at the other end of the spectrum. The agent has enough information to act autonomously and does so, but logs the action for a human to review after the fact. This is best suited to decisions that are easy to reverse, where real-time approval would add friction without adding much safety.
These three tools map fairly cleanly to three real-world scenarios: routine data capture with human sign-off, complex decisions requiring human judgement, and low-risk autonomous actions with an audit trail.
The Agent Feed Has Been Rebuilt
To support all of this, Microsoft has completely redesigned the agent feed inside model-driven apps. The previous version was essentially a notification panel tied to Copilot Studio activity. The new one is a shared workspace where humans and agents operate side by side.
The feed has two tabs. “Needs Attention” shows tasks waiting for human input, created by either invoke_data_entry or request_assistance. “Completed” shows everything else, including tasks the agent finished autonomously and logged via log_for_review.
For data entry tasks, the review experience is a side-by-side view: the original source document on the left, the agent-populated form on the right. You can check the agent’s work against the raw input, make corrections, and approve. When you correct the agent’s suggestions, those corrections feed back to improve future suggestions. The tool gets more accurate the more it is used in your environment.
A single model-driven app supports up to 10 simultaneous agents.
What This Means for You
For app makers and developers, this removes a significant barrier. Previously, connecting an AI agent to a model-driven app meant building and maintaining custom integrations. With the MCP Server, you configure the agent in Copilot Studio, write instructions in natural language referencing your Dataverse columns by their logical names, and the plumbing is handled for you. Access to the agent feed is controlled by security roles, with System Administrator and System Customizer roles having access by default. You can create custom roles to extend access to other users as needed.
For business users, the experience is designed to feel like a natural part of the app rather than a separate AI tool. Agents handle the volume work, you handle the judgement calls. Your focus shifts from entering data to reviewing and approving it, which is a better use of most people’s time.
The State Farm example Microsoft cites is a useful illustration. Claims estimates arrive in different formats with varied attachments. An agent processes each one, extracts the key field values, and surfaces a review task. A human checks it and approves. What was a slow, error-prone manual process becomes a supervision task.
A Few Things Worth Knowing Before You Start
This is a preview feature, so production use is not recommended yet, and functionality may be limited or change before general availability. It is currently available in English only and is rolling out first to US early release cycle environments.
The invoke_data_entry tool currently supports single line of text (None format), whole number, and decimal column types. Choice fields, lookup fields, date/time columns, and others are not yet supported. That is a notable limitation for more complex forms, and one to watch as the feature matures.
The Power Apps MCP Server works only with model-driven apps. Canvas apps and the newer vibe apps are not supported. And it is only available through Copilot Studio as the agent platform.
On the Dataverse side, Dataverse itself can also act as an MCP server, giving AI tools like Claude, GitHub Copilot, and others direct access to your tables and records. If you are using agents built outside of Copilot Studio to access Dataverse MCP tools, note that billing for those calls starts on December 15, 2025, unless you hold Dynamics 365 Premium or Microsoft 365 Copilot User licences.
The Bigger Picture
Microsoft built this on the back of its earlier agentic data entry feature in Power Apps, which saw strong adoption and generated enough user feedback to inform a broader strategy. The MCP Server takes that proven core capability and makes it available as a general-purpose tool that any Copilot Studio agent can use.
The result is a model where your existing business applications become the interface for human-agent collaboration, rather than requiring staff to work across separate AI tools and business apps simultaneously. For organisations already running model-driven apps on Power Platform, this is worth paying close attention to as it moves toward general availability.