MCP tool overview

Important

Windows 365 for Agents is in public preview. The feature is under active development and might change before general availability.

Windows 365 for Agents exposes capabilities through two complementary surfaces that map to the agent session lifecycle:

  • Microsoft Graph APIs for administration and session management. IT admins, agent makers, and partner applications use these APIs to stand up and govern capacity.
  • Model Context Protocol (MCP) tools for in-session operation. AI agents invoke these tools during a live session. For screen sharing, a partner app invokes them on a human's behalf.

Together, these surfaces cover provisioning the pool, acquiring a Cloud PC, performing work, and observing or assisting as needed.

On the Microsoft Graph API side, the Computer-Create plane uses the W365A Graph API and the W365 admin portal. Through these surfaces, administrators and independent software vendors (ISVs) can:

  • Provision Cloud PC agent pools.
  • Configure policies and images.
  • Register trusted partner callers.
  • Scale pool counts.
  • Attach metering through MAC billing.

The Computer-Get plane is a small runtime control surface for partner applications:

  • A Checkout call reserves a Cloud PC and returns the session identity and connection URLs.
  • A Checkin call releases the session and returns the Cloud PC to the pool.

Checkout is idempotent, so retries don't allocate duplicate sessions.

After the partner application acquires a Cloud PC, agents use MCP tools to operate it. These tools follow the open Model Context Protocol, so any agent that supports the protocol can discover and invoke tools without custom integration.

Computer-Do is the primary plane for in-session work. The agent calls into a set of built-in tools to control the Cloud PC. These tools cover:

  • Desktop interaction, including mouse, keyboard, and screenshot capture.
  • Window management.
  • Command execution.

More browser automation and UI accessibility capabilities are available or planned, which expand what an agent can reliably do inside the session.

For human supervision, the partner application uses Computer-See to observe and optionally co-drive the same Cloud PC session that the agent is using. Screenshare supports starting and stopping a real-time share, handing mouse and keyboard control to a human, and returning control to the agent.

Media and device redirection flow over the remote desktop stack: Azure Virtual Desktop and Remote Desktop Protocol (RDP) with IC3 media. This stack provides a viewing and control experience comparable to a standard remote desktop session, while the agent continues to operate in the same session context.

Surface Plane Endpoints Called by Purpose
Graph API Computer-Create W365A Graph API and W365 admin portal IT admin or ISV Shape and maintain the pool.
Graph API Computer-Get sessions (Checkout) Partner application Reserve a Cloud PC.
Graph API Computer-Get {sessionId} (Checkin) Partner application Release the Cloud PC.
MCP Computer-Do MCP AI agent Operate the Cloud PC.
MCP Computer-See, Computer-TakeControl screenshare (Start, Stop, TakeControl, ReleaseControl) Partner app, on behalf of a human Observe and co-drive.

How they fit together

The two surfaces work in sequence, with a clear handoff between callers:

  1. Admins and agent makers use Computer-Create to provision the pool.
  2. The partner application calls Checkout on Computer-Get to reserve a Cloud PC for a specific piece of agent work.
  3. The AI agent drives the Cloud PC through the Windows 365 Computer-Use MCP tools. Most calls flow through this plane.
  4. When needed, the partner application opens Computer-See on behalf of a human to observe or take over.
  5. The partner application calls Checkin on Computer-Get to release the Cloud PC when the work is done.

Next steps