Human-agent collaboration is moving beyond the cloud. It’s coming to the edge. At its 2025 Build conference, Microsoft unveiled the Windows AI Foundry, a development platform that will bring model building, fine-tuning, and local deployment capabilities to Windows 11.
The company also announced native support for the Model Context Protocol (MCP), a framework that allows AI agents to interact directly with Windows apps.
Together, these updates position the operating system as a powerful platform for building intelligent, locally run AI experiences that are faster, more private, and less dependent on the cloud.
Windows AI Foundry
Microsoft describes its new Windows AI platform as an “evolution” of Windows Copilot Runtime, a library of developer APIs unveiled at last year’s Build designed to support custom on-device models on Windows machines, such as the Copilot+ PCs. The Windows AI Foundry provides the tools and infrastructure to develop, fine-tune, and deploy lightweight models directly on the device. It’s laying the groundwork for AI agents to run natively on Windows 11.

According to Microsoft, the new Foundry has three key capabilities:
Support for Windows ML
Windows Machine Learning (ML) is an API used to deploy hardware-accelerated ML inferences locally. It is included with Windows 11 and allows developers to import their own AI models and deploy them on any device, regardless of the chip inside. The company touts that Windows ML simplifies deployment and automatically makes apps AI hardware future-proof.
To help get developers ready to launch their apps, Microsoft is updating its AI Toolkit for VS Code. The newest tools, available today in preview, support model conversion from PyTorch to ONNX, quantization (making models smaller and faster), compilation, and performance profiling. Developers should be able to use them to streamline the process of preparing and deploying AI models on Windows devices without needing complex builds or manual hardware tuning.
Microsoft discloses that Adobe, McAfee, Topaz Labs, Powder, and Wondershare are some of the developers already using Windows ML to deploy their models across the Microsoft silicon ecosystem.
Integration With Foundry Local and Other Model Catalogs
Developers can now access more ready-to-use open-source models, pulling from Foundry Local, Ollama, and Nvidia’s NIM, running on Windows silicon devices. These tie-ins create new opportunities for developers to see the best models for their AI apps running on local machines. Microsoft states Foundry Local will automatically detect the device’s hardware and recommend compatible models for developers. Moreover, developers can use the command line interface to browse, test, and interact with models running on a local server.
New Windows APIs for on-Device Fine-Tuning and RAG
The Windows AI Foundry comes with APIs powered by the same Windows inbox models as those on Copilot+ PCs. However, Microsoft is adding more capabilities, starting with LoRA (low-rank-adaptation) for the company’s Phi Silica lightweight model. Introduced at last year’s Build conference, it’s the smallest of the Phi models with 3.3 billion parameters. At the time, a spokesperson told me it was Windows’ inaugural locally deployed language model and was initially optimized to run on Copilot+ PC NPUs.
With the LoRA, developers can use custom data to fine-tune Phi Silica to their liking by updating only “a small subset” of its parameters. Microsoft claims this results in better performance for desired tasks without compromising Phi Silica’s overall abilities. LoRA for Phi Silica is only available in preview on Snapdragon X Series NPUs, but will expand to Intel and AMD Copilot+ PCs sometime in the future. It can be accessed in the Windows App SDK 1.8 Experimental 2.
Microsoft is also launching new APIs for Semantic search and Knowledge retrieval for LLMs. Developers can use them to implement semantic and lexical search in their apps. Moreover, they can support Retrieval-Augmented Generation. The APIs are available in a private preview today and will run locally on all devices.
MCP for Windows 11

Anthropic’s Model Context Protocol is gaining widespread adoption. This standard enables agents to connect to applications where data lives. Microsoft is no stranger to MCP, adding it to Copilot Studio in March. The company is now bringing it to edge devices with native support on Windows 11, though it’s only available as part of a private developer preview. The first group of partners to bring their agents to the operating system will be named in the coming months.
There are two parts to having MCP on Windows:
- MCP Registry on Windows: Agents will be able to tap into installed MCP servers, giving them access to app-specific functions and system context to deliver more useful, intelligent responses.
- MCP Servers for Windows: Speaking of MCP servers, this component includes a host of Windows system functionalities with which agents can interact. This means the bots could control your computer through a computer-use agent or an ad hoc scenario (IT troubleshooting).
Introducing App Actions
Another feature being announced is App Actions on Windows, a capability that helps developers build actions for specific features in their AI apps. Some examples might be having an agent create a new note in Microsoft OneNote, send an email to someone, schedule a meeting, open a recent document within File Explorer, search PDFs for a specific keyword, or clear the computer’s cache.
Developers can build and test actions using two tools:
- Microsoft’s App Action APIs, which enable action creation and are currently available in Windows SDK 10.0.26100.4188 or greater, and
- App Actions Testing Playground, a dedicated development environment used to test the functionality and user experience of these actions. This tool is available in the Microsoft Store.
Zom, Filmora, Goodnotes, Todoist, Raycast, Pieces for Developers, and Spark Mail are named among the first group of developers to start using App Actions.
Subscribe to The AI Economy
Extending AI PC Beyond Copilot+ PCs
It’s been a year since Microsoft introduced its take on the AI PC, and it has faced some stumbling blocks getting out of the gate. While the laptops appeared innovative, the company’s Recall feature—one of the more recognizable capabilities—faced heavy criticism and was removed before launch. It was finally added back to Copilot+ PCs last month.
Bringing AI agents to the edge would make it more valuable to have a Copilot+ PC, especially for knowledge workers. The devices are already equipped with state-of-the-art hardware explicitly optimized for AI tasks. However, the above features also make agents more accessible to the broader community of PC users. It opens up new opportunities for developers to create intelligent applications that can leverage the power of local hardware. By bringing MCP and App Actions to the broader Windows ecosystem, Microsoft is making it easier for a wider range of users to benefit from these AI-powered experiences.
AI agents on Windows shouldn’t be surprising. The company telegraphed its move earlier this month, announcing a bot that could change the operating system’s settings. And, of course, Microsoft’s Copilot agent is also on Windows.
Companies may find these updates interesting because they won’t have to buy expensive laptops. Additionally, they can develop specialized devices for their teams, such as salespeople, to help them better perform their jobs.
Alternatively, for the individual consumer, these new Windows 11 AI capabilities could raise more awareness of how the technology can enhance their lives.
Featured Image: Microsoft Corporate Vice President for Windows and Devices Pavan Davuluri on stage at the company’s Copilot+ PC announcement event on May 20, 2024. Credit: Ken Yeung
Subscribe to “The AI Economy”
Exploring AI’s impact on business, work, society, and technology.
Leave a Reply
You must be logged in to post a comment.