VS Code AI Update: Offline Support and Local Models Released

In the rapidly evolving landscape of developer tooling, local autonomy has emerged as the ultimate engineering battleground. On May 28, 2026, Microsoft officially shipped Visual Studio Code 1.122, a release that represents a monumental paradigm shift for privacy-first developers, enterprise architects, and engineering teams operating in secure, high-compliance environments. The headlining advancement in this iteration is the complete decoupling of the “Bring Your Own Key” (BYOK) AI features from the long-standing GitHub authentication requirement. By liberating the VS Code AI framework from cloud sign-in constraints, Microsoft has transformed its ubiquitous code editor into an unmatched local sandbox capable of running fully offline, air-gapped AI workflows.

The End of Mandatory Authentication: How VS Code AI Liberates the Offline Sandbox

Prior to version 1.122, developers wishing to use custom large language models (LLMs) or local models inside VS Code were met with a frustrating architectural paradox. Even if an engineer had set up a completely self-hosted inference server—such as Ollama or LM Studio running state-of-the-art open-weights models like DeepSeek, Llama, or Phi-4-mini—the editor still demanded an active GitHub authentication handshake before enabling the chat workspace. This telemetry-linked friction effectively locked out developers working in ultra-secure corporate intranets, defense systems, and medical laboratories where external network traffic and public cloud access are strictly prohibited by compliance protocols.

With VS Code 1.122, this authentication barrier is completely dismantled. The editor now supports the execution of chat assistants, multi-agent pipelines, and tool callouts entirely offline, without sending telemetry or requiring a Microsoft or GitHub cloud session. This update fundamentally changes the core philosophy of VS Code’s integrated intelligence, allowing teams to fully leverage local computing power for code generation, architectural analysis, and debugging while maintaining absolute control over their intellectual property.

How to Configure Your Decoupled VS Code AI Workspace

To establish a fully offline or private AI environment, VS Code 1.122 provides a streamlined configuration pipeline directly within the editor’s command interface. Setting up a BYOK provider silences all subsequent GitHub sign-in prompts and immediately unlocks the Chat panel. The setup process is straightforward:

1. Open the Command Palette by pressing Ctrl+Shift+P (Windows/Linux) or Cmd+Shift+P (macOS).
2. Search for and execute the Chat: Manage Language Models command.
3. Within the Language Models editor interface, select your desired model provider. The native system supports a wide range of external and local engines, including Anthropic, Azure AI Foundry, Gemini, OpenAI, OpenRouter, or any custom compatible endpoint.
4. To integrate local setups like Ollama or LM Studio, configure a custom OpenAI-compatible endpoint pointing to your local address (e.g., http://localhost:11434 for Ollama).
5. Once at least one custom BYOK provider is verified and enabled, the primary Chat workspace becomes available, and the system permanently suppresses cloud authentication prompts.

This decentralized architecture relies on a locally stored chatLanguageModels.json configuration file to save API routes, token contexts, and model metadata, ensuring your entire configuration remains locally auditable and reproducible across massive developer workstations.

Architectural Boundaries: Chat, Agents, and MCP vs. Inline Autocomplete

While this update is a massive victory for digital sovereignty, developers must understand the technical boundaries between the fully decentralized VS Code AI environment and features that still rely on GitHub infrastructure. Microsoft has partitioned the editor’s AI capabilities into distinct lanes:

• Fully Offline & Decoupled (No Sign-In Required): The primary Copilot Chat interface, developer-defined custom agents, custom utility tools, and active Model Context Protocol (MCP) servers. This allows users to write prompts, inspect local codebases, execute local agent workflows, and leverage complex tool-calling models entirely within an isolated network.
• Cloud-Tethered (GitHub Sign-In Required): Inline autocomplete suggestions, Next Edit Suggestions (NES), semantic code search, and features reliant on cloud-generated vector embeddings.

The reasoning behind this division is structural. High-speed inline autocompletions demand sub-100ms latencies and complex predictive algorithms that remain tightly coupled to GitHub’s proprietary cloud-completion infrastructure. However, for deep reasoning, systemic code refactoring, and multi-file analysis, local LLMs operating through the newly freed Chat interface are more than capable of handling heavy developer workloads.

A Financial Pivot: GitHub Copilot Shifts to Usage-Based Billing

Coinciding with the decoupling of local AI setups, the VS Code 1.122 release marks a major financial transition for developers utilizing official cloud extensions. GitHub Copilot has officially transitioned to a usage-based billing model. Moving away from flat-rate monthly subscriptions, this system calculates consumption via AI credits that are spent based on the complexity of developer interactions.

To help developers navigate this transition without running into unexpected costs, Microsoft has introduced several helper utilities inside the IDE:

• Model Picker Cost Indicators: The model selection dropdown now displays real-time pricing indicators based on input, output, and cached token costs for different models. Choosing lighter models for basic tasks helps developers preserve their credit pool.
• Updated Copilot Status Dashboard: An integrated billing HUD displays aggregate credit consumption, offering immediate visibility into how much budget an agent session or a complex chat prompt has consumed.
• Language Models Editor Upgrades: A centralized dashboard where developers can view precise model capabilities, maximum context window constraints, and distinct billing rates for active models.

Unleashing 1M Context Windows, Browser Emulation, and OpenTelemetry

Beyond the local AI revolution, VS Code 1.122 packs powerful modern capabilities designed to accelerate the development of complex web applications and distributed agent frameworks. Key technical upgrades in this release include:

Massive 1-Million-Token Context Support

For developers utilizing cloud-hosted endpoints such as Anthropic’s Claude or OpenAI’s GPT models, VS Code 1.122 now natively supports 1-million-token context windows. This massive upgrade allows developers to feed entire repositories, thousands of lines of documentation, or massive multi-file codebases directly into a chat prompt, dramatically reducing context fragmentation and enabling hyper-accurate, project-wide refactoring passes.

Integrated Browser Device Emulation

Web developers can now test the responsive behavior of web applications natively within the IDE’s integrated browser. By selecting the “Show Emulation Toolbar” from the browser’s overflow menu, developers can emulate screen dimensions, touch interfaces, mobile viewports, and custom user-agent headers. Furthermore, automated agents can programmatically trigger these emulations using Playwright scripts to catch UI responsiveness bugs during background test execution.

Granular OpenTelemetry Logging for Agents

To facilitate corporate auditing and agent monitoring, local agent sessions in VS Code 1.122 now emit rich OpenTelemetry signals. Emitted under the canonical github.copilot.* attribute namespace, these signals provide structured data on repository context, agent execution tracks, detailed tool parameters, and hook outcomes. This provides organizations with a robust audit trail to analyze AI behavior and ensure compliance without sacrificing developer productivity.

Linux Wayland Compatibility and the 1.122.1 Prompt Patch

Following the main release, the VS Code team quickly rolled out a critical dot-one patch (v1.122.1) to address several early-adopter issues. Most notably, the update resolves a highly frustrating bug affecting Linux developers running modern KDE Plasma 6 Wayland desktop environments. In the initial 1.122 builds, an aggressive screen-sharing permission loop would endlessly prompt users during startup, halting development. The v1.122.1 patch cleanly breaks this permission cycle, restoring smooth performance and system stability across Linux, Windows, and macOS distributions.

Ultimately, Visual Studio Code 1.122 stands as a monumental milestone for the developer ecosystem. By removing the GitHub sign-in requirement for BYOK models, Microsoft has acknowledged the growing demand for absolute data privacy and offline capability. Whether you are running a fully air-gapped system powered by Ollama and DeepSeek, or orchestrating multi-agent networks over local MCP servers, the latest VS Code release ensures that your AI assistant remains entirely on your terms.