Who should use Foundry Toolkit
Foundry Toolkit supports different roles and levels of AI development experience:- Application developers can add generative AI features to web, desktop, and mobile apps.
- AI and machine learning engineers can compare, customize, optimize, and deploy models and agents.
- Data scientists and researchers can experiment with prompts, models, datasets, and evaluation criteria.
- Educators and students can explore generative AI concepts through interactive playgrounds and local models.
How the Toolkit workspace is organized
Foundry Toolkit separates resources that you have from actions that you can take. The extension has three main sections.
The exact tools and resource types can change as the extension evolves. Use the
Toolkit view to see the features in your installed version.
Work with models
Foundry Toolkit provides an end-to-end model development workflow:- Use the Model Catalog to discover models from Foundry and external providers, or use local models through ONNX, Ollama, and Foundry Local.
- Compare models and test prompts, parameters, images, and attachments in the Model Playground.
- Convert, quantize, optimize, and evaluate supported models for local Windows deployment.
- Fine-tune supported models locally with a GPU or remotely with Azure Container Apps.
- Profile CPU, GPU, and NPU usage for supported Windows machine learning workloads.
Build and operate agents
Foundry Toolkit supports prompt agents and code-based hosted agents. Start with Create an agent to compare Agent Builder, hosted-agent samples, and Copilot-assisted coding.
Agent Builder also supports locally stored prompts. Their tools, structured
output, and dataset evaluation options differ from those of Foundry prompt
agents. Local storage doesn’t mean that model inference runs on your machine.
See Work with local prompts.
During development, you can:
- Use Agent Builder to test conversations, save prompt-agent versions, and generate client code that calls a saved Foundry agent.
- Use the Tool Catalog to discover Foundry tools, Model Context Protocol (MCP) servers, and toolboxes, and then add them to agents.
- Use Agent Inspector to debug local agents, inspect streaming responses and tool calls, and visualize workflow execution.
- Deploy hosted agents to Foundry Agent Service from source code or a container image.
- Test deployed agents in the Hosted Agent Playground, inspect logs and traces, and manage versions.
- Evaluate models, prompts, and agents with datasets, built-in evaluators, or custom criteria.
Declarative workflows in Microsoft Foundry retire on December 1, 2026.
Use Microsoft Agent Framework for new workflow development. This retirement
doesn’t affect code-based orchestration in hosted agents. See the
workflow migration guide.
Manage Foundry resources
Sign in to Azure and set a Foundry project to manage cloud resources from Visual Studio Code. You can:- Browse Foundry projects and the resources available to your account.
- Discover and deploy models from the Foundry model catalog.
- View model endpoints and authentication information.
- Create and version prompt agents, and deploy and test hosted agents.
- Inspect and test existing declarative workflows before migration.
- Browse tools and knowledge sources used by your agents.
- Open evaluations, conversations, logs, and traces during development.