Microsoft Rolling Out Copilot Local PC File Access, The Register Reports
Microsoft brings local AI agents to Windows file systems
Microsoft is rolling out hybrid intelligence features that grant its Copilot assistant access to files on personal computers, according to a report published by The Register. The functionality combines on-device models with cloud-based processing.
The Tech TL;DR:
- Copilot gains access to local file systems through agentic search and productivity integrations in Microsoft Office.
- Execution Containers (MXCs) sandboxes run local agents with runtime permissions, using Microsoft’s HydraFusion router to direct simple tasks locally and complex queries to cloud models like GPT-6 or Claude.
- Support for Windows ML is extending to Llama.cpp to utilize various hardware accelerators, rolling out to GitHub Copilot next week and the Copilot app over the coming months.
HydraFusion routes tasks between local and cloud models
The system relies on Microsoft’s HydraFusion model router to decide whether an AI task runs locally or on cloud servers. Windows may execute simple or privacy-sensitive tasks using a local model, while routing complex or performance-sensitive requests to cloud-based models such as OpenAI’s GPT-6 or Anthropic’s Claude.
To secure local processing, Microsoft runs local agents inside Microsoft Execution Containers (MXCs). These are agent sandboxes designed with fine-grained permissions applied at runtime. According to Microsoft EVP of Windows and Devices Pavan Davuluri, these containers make Windows the most secure platform for agents. The infrastructure is supported by extending Windows ML—an inference framework built on the ONNX runtime—to Llama.cpp, enabling inference engines to use a broader range of hardware accelerators.
Office integration and natural language search capabilities
Copilot’s agentic functionality interacts directly with Microsoft Office applications. In a demonstration, the assistant located and organized a user’s tax documents before drafting an email with the files attached and addressed to an accountant. Microsoft states that the assistant will only search documents when explicitly requested by the user.
Beyond file access, the operating system is receiving AI-enhanced search features that allow users to surface settings like dark mode using natural language. The interface also supports conversational commands for tasks such as sending text messages. Microsoft is also introducing an onboarding experience to assist users deploying agent harnesses like OpenClaw.
Local model processing affects battery life and performance
The hybrid setup builds on the architecture introduced with Copilot+ PCs in 2024 alongside Qualcomm Snapdragon X Elite SoCs, and is announced alongside a new Nvidia N1X-powered Surface Laptop. Keeping low-priority tasks on the device reduces server load for Microsoft and addresses privacy concerns raised by continuous cloud data transmission.
However, running large language models locally introduces trade-offs for client hardware. The processing overhead impacts battery life, generates fan noise, and creates workloads that may affect overall system responsiveness for users relying on mobile devices.