Foundry Toolkit for VS Code

Provides a unified model discovery interface within VS Code that aggregates models from 8+ sources (Microsoft Foundry, GitHub Models, OpenAI, Anthropic, Google, NVIDIA NIM, Ollama, ONNX) with side-by-side comparison capabilities. The extension maintains a tree view in the sidebar with a 'Model Catalog' section that dynamically populates available models based on configured API keys and local installations, enabling developers to evaluate and select models without leaving the editor.

Provides an embedded chat interface within VS Code for real-time model testing and prompt experimentation. The playground supports multi-modal inputs (text, images, attachments), parameter tuning (temperature, top-p, max tokens), and streaming response visualization. Developers can test prompts against any model in the catalog without leaving the editor, with full parameter control and response inspection.

Enables agents to route requests to multiple models simultaneously or sequentially, compare outputs, and select the best response based on custom criteria. The extension provides orchestration patterns (parallel execution, fallback chains, ensemble voting) and comparison metrics (similarity, relevance, cost) to help developers optimize agent behavior. Results from all models are captured and compared in the debugger.

Manages agent deployment to Microsoft Foundry and other hosting environments, including versioning, rollback, and environment configuration. Developers can deploy agents directly from VS Code, manage multiple versions, configure environment-specific settings (API keys, model selections), and monitor deployed agent health. The extension handles deployment packaging and orchestrates the deployment process.

Provides dual-mode agent development: a no-code prompt-based agent builder for simple workflows and a code-based hosted agent framework for complex multi-step agents. Both modes support structured output generation (JSON schemas, typed responses) and integrate with the debugger for real-time execution visualization. The builder abstracts away boilerplate agent scaffolding while maintaining full code access for advanced customization.

Provides F5-based debugger integration for agent execution with real-time streaming response visualization and multi-agent workflow inspection. When launching an agent with F5, the extension captures execution traces, tool calls, and model responses, displaying them in a structured timeline view within VS Code. Developers can inspect intermediate states, tool invocations, and response generation without external logging or debugging tools.

Enables systematic model evaluation against datasets using a combination of built-in evaluators (F1 score, relevance, similarity, coherence) and custom evaluation criteria. Developers upload or reference datasets, define evaluation metrics, and run batch evaluations across models to compare performance. Results are displayed in a structured comparison view with metrics aggregation and per-sample analysis.

Enables fine-tuning of models on local GPU hardware or via Azure Container Apps for cloud-based training. The extension abstracts away training infrastructure setup, handling data preparation, training loop orchestration, and model checkpointing. Developers specify a dataset, select a base model, configure training parameters (learning rate, epochs, batch size), and launch training either locally or in the cloud with progress monitoring within VS Code.

Provides automated model conversion and optimization workflows for transforming models between formats (Hugging Face to ONNX, quantization for edge deployment). The extension integrates with Hugging Face model hub, applies quantization techniques (int8, int4, or other precision reductions), and generates optimized models ready for local deployment via ONNX runtime or Ollama. Conversion progress and optimization metrics are displayed within VS Code.

Collects and visualizes performance metrics during agent execution, including latency per step, token usage, API call costs, and resource consumption. Traces are captured automatically during F5 debugging or explicit trace collection, aggregated into a timeline view, and exported for analysis. Developers can identify bottlenecks, optimize expensive operations, and track cost implications of agent design choices.

Provides CPU/GPU/NPU resource usage diagnostics and execution provider analysis for ONNX models running on Windows. The profiler captures Windows ML event traces, analyzes execution provider selection (CPU, GPU, TensorRT, CoreML), and reports resource consumption (memory, compute utilization). Results are displayed in VS Code with per-operation breakdown and optimization recommendations.

Enables agents to invoke external tools via Model Context Protocol (MCP) integration, allowing structured function calling with schema-based tool definitions. Developers define tools as MCP resources, agents discover and invoke them with type-safe parameters, and results are returned to the agent for further processing. The extension manages tool registration, parameter validation, and error handling.