Build AI Apps with Azure AI Projects SDK

Azure AI Projects Python SDK (azure-ai-projects) builds AI applications on Microsoft Foundry, giving one client (AIProjectClient) access to seven operation areas - agents (CRUD, versions, threads, runs), connections, deployments, datasets, indexes, evaluations, and red-team operations. It offers two ways to call a deployed model: native Foundry operations through client.agents.create_agent(), or an OpenAI-compatible client obtained via client.get_openai_client() for standard chat.completions.create() calls. The full agent-conversation loop follows a thread/message/run pattern:

### 1. Create thread
thread = client.agents.threads.create()

### 2. Add message
client.agents.messages.create(
    thread_id=thread.id,
    role="user",
    content="What's the weather like?",
)

### 3. Create and process run
run = client.agents.runs.create_and_process(
    thread_id=thread.id,
    agent_id=agent.id,
)

### 4. Get response
if run.status == "completed":
    messages = client.agents.messages.list(thread_id=thread.id)
    for msg in messages:
        if msg.role == "assistant":
            print(msg.content[0].text.value)

Agents can be given tools - CodeInterpreterTool, FileSearchTool for RAG over uploaded documents, BingGroundingTool, AzureAISearchTool, FunctionTool, OpenApiTool, McpTool for MCP servers, MemorySearchTool, and SharepointGroundingTool - and can be versioned via PromptAgentDefinition and create_version() for production deployment rather than left as a single mutable agent.

Use this SDK for full Foundry-integrated projects that need connections, model deployments, datasets and indexes, and evaluation alongside agents - it explicitly contrasts itself with the lower-level azure-ai-agents SDK, which lacks versioning, connections, deployments, dataset/index management, and evaluation support and is meant for standalone agent apps instead. Use the async client (azure.ai.projects.aio, with async with) for real-time or concurrent workloads, and streaming rather than create_and_process when you need a real-time UX. It is not the right choice for a standalone agent that doesn't need Foundry's project-level features - connections, deployments, datasets, evaluation - use azure-ai-agents for that instead.

Input is model deployment names, agent instructions and tools, thread messages, and evaluation datasets; output is created and versioned agents, thread runs with assistant messages, connection and deployment listings, dataset and index references, and evaluation runs scored against built-in criteria such as fluency and task adherence via the OpenAI-compatible client's evals.runs.create(). Memory stores can be created and attached to an agent's tool_resources for persistent conversation memory across runs.

Authenticates via DefaultAzureCredential; integrates with Azure AI Search, Bing Grounding, SharePoint, and MCP servers as agent tools, and exposes a full OpenAI-compatible client for using standard OpenAI SDK calls against the same deployed model. The reference documentation covers all 373 SDK exports (v2.0.0b4), and a bundled scripts/run_batch_evaluation.py provides a CLI for batch evaluations.

Python developers building full Foundry-integrated AI applications - agents plus connections, deployments, datasets, indexes, and evaluation - who need the higher-level AIProjectClient rather than the standalone, lower-level azure-ai-agents SDK.