Microsoft turns Windows into an AI agent playground at Build 2025

Running AI models locally on Windows has long been a fragmented mess of shipping bloated runtimes, wrestling with driver variations, and praying an end user's NPU would actually cooperate.

At Build 2025 on May 19, Microsoft tackled that friction head-on. The company overhauled its desktop developer platform, officially rebranding Windows Copilot Runtime into Windows AI Foundry, baking in native support for Anthropic’s Model Context Protocol (MCP), and open-sourcing the Windows Subsystem for Linux (WSL).

Windows ML strips out the deployment headaches

The most practical upgrade for builders is Windows ML, the built-in inference runtime that evolves DirectML.

Instead of forcing developers to bundle proprietary machine learning runtimes, drivers, and execution providers into their app installers, Windows ML now handles that layer inside the operating system. It detects whether the machine is running AMD, Intel, NVIDIA, or Qualcomm silicon, then pulls and configures the exact execution provider required. As hardware upgrades roll out, the underlying system handles compatibility without breaking existing applications.

To simplify sourcing models, Windows AI Foundry directly integrates Foundry Local alongside model catalogs like Ollama and NVIDIA NIMs. In preview, developers can run a single command (winget install Microsoft.FoundryLocal) to browse compatible models tailored to their hardware.

For on-device intelligence on Copilot+ PCs, Microsoft is also introducing LoRA fine-tuning for its built-in Phi Silica small language model (SLM), alongside new system-level APIs for semantic search and RAG workflows.

Standardizing agents across the desktop

Beyond basic model inference, Microsoft is laying the plumbing for desktop agents by natively supporting MCP inside Windows 11.

This open standard allows local AI agents to connect directly with native Windows apps via standardized hooks. Combined with a new capability called App Actions, software can expose specific features to system agents, allowing autonomous assistants to perform tasks across multiple programs on the PC. Microsoft said a private developer preview will roll out to partners in the coming months.

On the hardware front, PC makers are lining up heavy metal to run these local workloads. Dell unveiled the Pro Max Tower T2 desktop and Pro Max 16 laptop featuring NVIDIA RTX PRO Blackwell generation GPUs, slated for release in July 2025. HP also introduced its compact ZBook Ultra G1a powered by AMD's Ryzen AI Max PRO.

Why it matters

Microsoft is making a calculated push to keep Windows indispensable as the operating system for local AI development.

For developers, offloading runtime packaging and hardware acceleration to Windows eliminates massive operational drag. Teams at Adobe, McAfee, and Topaz Labs are already on board. Running inference and fine-tuning locally not only slashes recurring cloud API costs, but it also gives privacy-conscious software a clean path to function offline.

Industry-wide, Microsoft’s endorsement of Anthropic’s MCP standard across Windows, Azure, and GitHub is a significant signal. Rather than erecting a proprietary, walled-off agent protocol inside Windows, Microsoft appears willing to back an emerging cross-platform standard to accelerate ecosystem adoption.

We think the execution strategy is smart, but it comes with real risks. Abstracting the silicon layer through Windows ML relies heavily on chipmakers keeping their execution providers flawlessly updated. If an Intel or Qualcomm driver update introduces regressions, desktop app developers will be the ones fielding user bug reports. Microsoft has built an ambitious foundation, but holding disparate silicon partners to high software quality will be the real test.