During Build 2025, Microsoft first announced that native MCP (Model Context Protocol) support was coming soon to Windows. Today, at Ignite 2025, Microsoft announced the public preview of native support for Model Context Protocol (MCP) on Windows, using which AI agents can connect with apps and tools to perform tasks on a user"s behalf.
Microsoft today also announced the public preview of Windows On-Device Registry (ODR), a manageable repository of agent connectors, which are just MCP servers. Microsoft will have several built-in agent connectors, but Windows will also support 3rd party local and remote agent connectors. Once these agent connectors are registered with the On-Device Registry, agents can connect to take advantage of the new skills to complete tasks for users.
The Redmond giant also introduced the public preview of two built-in agent connectors for File Explorer and System Settings.
- File Explorer connector: This will enable secure access to local user files on the system so agents can act more efficiently and perform desired file operations, all with user consent. Agents will be able to use this connector to perform a wide range of operations on local files on users’ devices, including searching, reading, writing, and editing. The file search capabilities will support natural language queries, allowing agents to find files based on descriptions, content, metadata, and, for images, enhanced search based on image classification.
- Windows Settings connector: This will help agents handle complex setting operations by modifying a user’s device settings such as Bluetooth, network, and more.
Finally, the company unveiled the private preview of Agent workspace, an isolated, policy-controlled, and auditable environment in which agents can interact with other apps on Windows without affecting the user’s primary session. Microsoft mentioned that all agentic interactions involving MCP and computer-using agents will run in Agent workspace. This will allow organizations" IT team to easily distinguish between agent and user actions.