Manage Azure AI Projects with TypeScript

This skill uses the Azure AI Projects SDK for TypeScript (@azure/ai-projects plus @azure/identity), a high-level client for Azure AI Foundry projects covering agents, connections, model deployments, datasets, search indexes, evaluators, and agent memory stores. An AIProjectClient is constructed from a project endpoint and a DefaultAzureCredential, and exposes seven operation groups: agents, connections, deployments, datasets, indexes, evaluators, and memoryStores. A key pattern is calling client.getOpenAIClient() to get an OpenAI-compatible client for responses.create (single-turn generation) and conversations.create (multi-turn threads) - the recommended path for actually running models, files, and vector stores. Agents are created and versioned via client.agents.createVersion(name, {kind: "prompt", model, instructions, tools}), with tool configurations shown for Code Interpreter (code_interpreter), File Search (file_search against a vector store), Web Search (web_search_preview with an approximate user location), Azure AI Search (azure_ai_search against a project connection and index), custom Function tools (JSON-schema parameters with strict: true), and MCP tools (server_label, server_url, require_approval). Running an agent means creating a conversation via the OpenAI client, then calling responses.create with an agent: {name, type: "agent_reference"} body referencing the versioned agent, followed by explicit cleanup (conversations.delete, agents.deleteVersion). Connections can be listed, fetched by name, fetched with credentials (getWithCredentials), or resolved to a type's default (getDefault("AzureOpenAI", true)) so credentials come from the project rather than being hardcoded. Deployments can be listed and filtered by modelPublisher. Datasets support uploading a single file or a whole folder with a name/version, fetching, listing versions, and deleting. Indexes are created/updated from an AzureAISearchIndex config (name, connection, index name), listed, and deleted.

Use this skill when building a TypeScript application on Azure AI Foundry that needs to create and run versioned AI agents with tools (code execution, file/web/Azure AI Search, custom functions, MCP servers), manage project connections and model deployments, or upload datasets and search indexes for grounding. It is not needed for a bare OpenAI API integration with no Azure AI Foundry project structure - in that case the underlying OpenAI SDK alone (obtained here via getOpenAIClient()) would suffice without the project/agent/connection layer.

Input is agent configuration (model, instructions, tools), conversation messages, dataset files, or index definitions. Output is a versioned agent reference usable in conversations, model responses/conversation objects, uploaded dataset/version records, and created/updated search indexes.

Built on Azure AI Foundry via @azure/ai-projects, authenticated with Azure Identity's DefaultAzureCredential, wrapping an OpenAI-compatible client for responses/conversations, Azure AI Search for the search-index and Azure AI Search tool integrations, and optional @azure/monitor-opentelemetry tracing.

npm install @azure/ai-projects @azure/identity

TypeScript developers building agentic applications on Azure AI Foundry who need versioned agents with tool integrations (code, search, functions, MCP), plus project-managed connections, deployments, datasets, and search indexes, without hardcoding credentials.