Microsoft Foundry Toolkit for VS Code Goes GA: Full AI Development Lifecycle in Your Editor

Microsoft has officially released the Foundry Toolkit for Visual Studio Code as a generally available product, marking a significant milestone in AI development tooling. The toolkit consolidates what were previously separate extensions into a unified experience that supports the entire AI development lifecycle—from initial model exploration through production deployment and edge optimization.

What's New in GA

The GA release brings several key improvements and consolidations:

• Extension consolidation: Multiple previous extensions have been merged into a single, streamlined toolkit
• Model catalog expansion: Access to 100+ models from GitHub, Microsoft Foundry, OpenAI, Anthropic, Google, plus local models via ONNX, Foundry Local, or Ollama
• Rebranding: Unified under the Microsoft Foundry brand
• Enhanced workflows: Improved experimentation, agent development, and evaluation capabilities

Model Experimentation Made Simple

The toolkit's Model Playground is designed for rapid experimentation. Users can compare two models side-by-side, attach files for multimodal testing, enable web search, adjust system prompts, and watch streaming responses in real-time. When experimentation yields promising results, the "View Code" feature generates ready-to-use code snippets in Python, JavaScript, C#, or Java—translating the tested API call directly into the user's preferred language.

This approach eliminates the traditional friction of moving from experimentation to implementation, allowing developers to go from curiosity to testing in minutes rather than hours.

Two Paths for Agent Development

Foundry Toolkit supports both rapid prototyping and professional development through two distinct but connected paths:

Path A: No-Code Agent Builder

The low-code interface allows users to:

• Define agent instructions and refine them with a built-in Prompt Optimizer
• Connect tools from the Tool Catalog, including local MCP servers
• Configure MCP tool approval settings (automatic or manual)
• Switch between agents instantly with auto-save functionality
• Export prototypes directly to code when ready for production

Path B: Professional Development

For teams building complex systems, the toolkit provides:

• Scaffolding for Microsoft Agent Framework, LangGraph, and other orchestration frameworks
• Agent Inspector for debugging with full VS Code debugger support
• Real-time workflow visualization and local tracing
• One-click deployment to Microsoft Foundry Agent Service
• Hosted Agent Playground for testing without leaving VS Code

The Bridge Between Paths

The toolkit includes seamless conversion between prototyping and professional development. Agent Builder prototypes can be exported directly to code with a single click, generating a project that includes instructions, tool configurations, and scaffolding. GitHub Copilot integration with the Microsoft Foundry Skill helps maintain momentum during the transition to code.

Integrated Evaluations

Quality measurement is built into the workflow at every stage. Users can define evaluations using pytest syntax, run them from VS Code Test Explorer alongside unit tests, and analyze results in a tabular view with Data Wrangler integration. For larger-scale testing, the same definitions can be submitted to run in Microsoft Foundry, making evaluations versioned, repeatable, and CI-friendly.

Edge AI Capabilities

The toolkit extends beyond cloud development to support local AI workloads on Windows devices:

Model Conversion

A complete pipeline converts models from Hugging Face to hardware-ready ONNX format, supporting quantization and evaluation. The conversion targets multiple Windows ML execution providers:

• MIGraphX (AMD)
• NvTensorRtRtx (NVIDIA)
• OpenVINO (Intel)
• QNN (Qualcomm)
• VitisAI (AMD)

Profiling Tools

Real-time visibility into CPU, GPU, NPU, and memory consumption helps optimize model performance. Three profiling modes cover different scenarios:

• Attach at startup
• Connect to running processes
• Profile ONNX models directly

Windows ML Event Breakdown provides detailed insights into execution phases, while operator-level tracing shows how graph nodes are dispatched across hardware.

Fine-Tuning

Domain-specific adaptation is possible through LoRA (Low-Rank Adaptation) training for Phi Silica. The workflow uses Azure Container Apps for cloud training, tracks loss curves, and provides cloud inference for validation before downloading adapters for edge deployment.

Getting Started

The toolkit is available through the VS Code Marketplace. Microsoft provides hands-on labs and samples through GitHub repositories:

• Foundry Toolkit Samples
• VS Code Lab

Users are encouraged to share feedback and file issues on GitHub, and to join the broader conversation in the Microsoft Foundry Community.

The GA release represents Microsoft's commitment to providing a comprehensive, integrated development environment for AI that spans the full lifecycle from experimentation to production deployment, whether in the cloud or at the edge.