AI news · October 9, 2026

Microsoft makes Execution Containers generally available with local coding models

agentswindowslocal-modelscoding

Microsoft's Oct 7 Windows release pairs generally available MXC isolation with a local 137B/6.8B-active coding model and 256K context, giving agent builders a path to keep routine work on-device.

MAI Code total parameters
137B
active parameters
6.8B
local context window
256K
Made With Models illustration for this story

Microsoft's Oct 7 Windows event shipped two pieces aimed at agent builders. Microsoft Execution Containers are generally available for Windows 11 and use a JSON policy plus SDK to limit an agent's files, network, UI, and process access across Windows, macOS, and Linux. The public GitHub code is still marked early preview and warns that current policies can be too permissive, so do not treat it as finished security software.

The same event described local MAI Code 1.1 Flash at 137B total parameters and 6.8B active, compressed to 3-bit precision for on-device use with a 256K context. Microsoft says local calls in GitHub Copilot have no per-inference charge on supported devices; support and hardware availability still matter. This is a product direction, not proof that a local model will match a hosted one on your workload.

What you can do with it

Try the MXC repository with a disposable coding task and start with a narrow file and network policy. If you have a supported device, try MAI Code through GitHub Copilot or VS Code. Treat the public MXC code as an early preview and review every allowed path.

Our take

The useful move is the combination: local coding plus a declared resource policy. The catch is maturity. Microsoft's own repository warns about permissive policies, so teams should use it to learn the integration pattern, not to claim that an agent is safe by default.

Source: Windows Experience Blog ↗ — Made With Models writes the brief; the reporting is theirs.