The promise of running AI coding assistants locally - no subscription fees, no data leaving your machine, no usage caps - just got more complicated. VS Code's new Agents window does let you point the editor at a locally-running AI model, but Microsoft kept the GitHubb Copilot](/tools/github-copilot/) subscription requirement in place. You still need an active Copilot plan and an internet connection to use it.
That defeats most of the point for developers who turn to local models specifically to avoid subscriptions and cloud dependencies.
What the Agents Window Actually Does
The Agents window is a new interface in VS Code that lets you configure and interact with AI coding assistants directly inside the editor. It supports connecting to locally-hosted models - meaning AI models you download and run on your own hardware rather than through a call to a cloud server. Tools like Ollama and LM Studio, which let you run open-source models entirely on your own machine, are the obvious targets here.
On paper, that's genuinely useful. Local models give you full control over your data, no per-query costs, and the ability to work in environments where outbound API calls aren't allowed. Developers at companies with strict data policies often prefer local setups for exactly these reasons.
The Strings Attached
Requiring a GitHub Copilot subscription - which starts at $10/month for individuals - alongside local model support creates a strange arrangement. You're paying for cloud AI access while running your own hardware and your own model. The internet connection requirement suggests VS Code is phoning home for authentication or telemetry even when the actual AI processing happens locally on your machine.
Microsoft hasn't explained publicly why local model support is gated behind a Copilot plan. The most likely explanation is that the Agents window framework itself is classified as a Copilot feature, with local models treated as an alternative inference backend rather than a standalone capability.
For developers who want genuinely self-contained local AI coding help, Cursor (which supports local models without a mandatory cloud subscription), Continue, and Aider remain cleaner options. VS Code's implementation looks like a half-step - the local model support is real, but the requirements attached make it less useful for the people most likely to want it.