The hardest problem in agent design is not making the agent work. It is deciding what happens when the agent should not act alone - when a record needs a human eye before it is committed, when a judgement call sits outside the agent's remit, when volume needs supervision rather than automation. Until recently the answer was custom plumbing. The Power Apps MCP Server is Microsoft's answer to making it standard plumbing.
What Microsoft says it is
The foundation is the Model Context Protocol. In Microsoft's words:
"The model context protocol (MCP) is an open protocol that enables seamless integration between large language model (LLM) applications and external data sources and tools. Your agent can use the Power Apps MCP Server to communicate with your Power Apps, providing right human-in-the-loop supervision or agentic workflows."
Microsoft describes two types of capability:
"The Power Apps MCP Server equips your agent with two types of capabilities: 1. Automate repetitive app tasks ... 2. Supervise agent activity: The Power Apps MCP Server also provides specialized tools to business users to supervise any agent activity in the agent feed. Agents can now handoff control to humans for review, assistance, and steering with the MCP tools."
Two facts worth noting before you build. First, this is a preview feature - Microsoft's standard caution applies: "Preview features aren't meant for production use and might have restricted functionality." Second, it is already becoming the standard path: "Starting May 1st, 2026, agent feed only supports agents that use the Power Apps MCP Server to create tasks."
Source for all of the above: Work with Power Apps MCP server, Microsoft Learn.
What it means in practice - the FAb view
Strip the protocol language away and the Power Apps MCP Server does one important thing: it gives agent work a place where humans already are. Agents create records, fill forms and raise tasks; people review, approve, correct and steer - all through a single agent feed inside a model-driven app, instead of a bolted-on approval flow or an inbox full of notifications.
That is the missing layer between autonomous processing and human oversight. Our shorthand for the design principle: the agent is responsible for the work, the person is accountable for the outcome.
We are working directly with the Microsoft product team on this preview, and one of our clients is already using it to build a high-volume contract review system - agents handle the volume of inbound contracts, people handle the judgement calls the feed surfaces to them. It is the pattern we expect most supervised-agent builds to follow from here.
Where it fits in an agent architecture
On this site we describe three shapes of agent solution - contextual assistants, supervised agents, and platform solutions that are not agents at all. The Power Apps MCP Server belongs squarely to the second: it is the supervision mechanism for process-driven, human-in-the-loop agents built on Power Apps, Copilot Studio and Dataverse. If you are choosing between those shapes, start with our agents page or the platform comparison in choosing the right Microsoft agent platform.
Because it is in preview, our advice is measured: build with it now if agent supervision is on your roadmap, design for the agent feed as the review surface, and keep production commitments until general availability. That is what we are doing ourselves.