Announced at Microsoft Ignite 2026

Agents that stay inside the lines.
Local models decide. Windows enforces.

Cephable agents are coming to Windows with support for Microsoft Execution Containers (MXC). They plan and act on local models on your PC, keep more of the task in working memory, and can only reach the files and networks your policy allows.

  • Public beta · early December
  • GA for Windows · January 2027
  • Nothing leaves the machine
The Cephable agent
local model · on this device
Cephable AI Workflows
Making a plan…
Working memorysized to this PC
Windows · MXC container
policy set by IT · enforced by the OS
Agent container policy
IN BOUNDS FOR THIS AGENT
  • Projects folderAllowed
  • DocumentsAllowed
  • Approved websitesAllowed
  • Finance shareBlocked
  • Any other networkBlocked
WHAT WINDOWS SAW
How it works

Three layers, one agent your security team can approve

The thinking stays on the device. The doing stays inside a boundary that Windows enforces. Neither depends on a cloud service.

Layer 1 The Cephable agent You describe the outcome. The agent plans the steps, works through your apps and files, and shows each step as it goes so you stay in the loop.
Layer 2 Local models Planning and language run on local models using the GPU, CPU and NPU in the machine. Prompts, files and screen content are never sent out for inference.
Layer 3 MXC containment Each agent runs in its own contained environment. Policy decides which files and networks are in bounds, and Windows applies it while the agent works.
Context window expansion

A bigger working memory, sized to your PC

Long tasks fail when an agent forgets what it already did. Cephable sizes the agent’s working memory to the graphics and system memory the machine actually has, so a run keeps more steps, results and instructions in view.

Thin and light Everyday tasks and short workflows
AI PC with NPU Full document and multi-app workflows
Workstation Long, many-step runs that stay on track

Illustrative. Working memory scales automatically on each device, with no cloud context service and no usage meter.

Everything else

Built for teams that have to say yes to security

Private, predictable and governed from the operating system up.

Private by designInference runs on the device. Prompts, files and screen content stay on the PC.
No token meterLicensed per user per year with no consumption fees, so agents can run all day without a bill that grows with use.
Policy from WindowsContainment is enforced by the operating system, which gives IT one control point instead of a setting inside every app.
Works offlineTurn the network off and the agent keeps working. The airplane-mode test is the quickest proof the work is local.
Connects to your toolsA launch partner for Model Context Protocol on Windows, and reachable over local HTTP from your own scripts and apps.
Every chip, tunedOptimized across Intel, Qualcomm, AMD, NVIDIA and Apple silicon to get the most from the hardware in the machine.
0token or consumption fees
1policy boundary, enforced by Windows
0bytes off this device for inference
Timeline

From Ignite to your desktop

November 17, 2026Announced at Microsoft Ignite alongside the Windows local AI partners.
Early December 2026Public beta for Windows. Agents with MXC support open to everyone who signs up.
January 2027General availability for Windows.

Microsoft, Windows and Microsoft Ignite are trademarks of the Microsoft group of companies. Dates are subject to change. Animations are illustrations.