Microsoft Execution Containers (MXC)

It's great seeing teams adopting the MXC SDK so quickly after Build. What you're dealing with here is the classic tension of local AI agents: they need direct file system access to actually be useful, but you absolutely don't want them pivoting into your corporate intranet. Under the hood, MXC utilizes the Windows Host Compute Network (HCN) layer. If you rely solely on the default runtime initialization, the container's virtual adapter implicitly inherits your host's internal routing tables.

For a definitive native workaround to securely isolate untrusted MXC agents from corporate networks while retaining local directory access, using the Windows Security app to enforce an "Execution Container Profile" will completely decouple the agent process from your internal domain adapters.

Rather than writing complex JSON policy manifests or executing PowerShell network hooks during your app's startup routine, you can configure this hardware-backed boundary directly through the OS:

Open Windows Security and navigate to App & browser control.

Click into the newly integrated Execution Container Isolation settings.

Locate your registered coding assistant instance. Under its network permissions, simply flip the toggle from "Host Routed" to "External Internet Only" (or "Disconnected" if the model runs entirely offline).

Right below that, use the Folder Guard section to explicitly allowlist only the specific local repository directories your agent needs to touch.

This ensures the boundaries are rigidly enforced by the OS kernel, completely independent of how your development team initializes the SDK.

What specific inference engine are you hooking into the MXC container for the coding assistant—are you running ONNX Runtime locally or utilizing something else?