# OpenAI Agents SDK TypeScript A TypeScript SDK for developers building text, sandbox, and realtime voice agents. - Findings24: - ウェブサイト: - 料金: Free - カテゴリー: AIエージェントビルダー, AIエージェント基盤, AI SDK ## OpenAI Agents SDK TypeScriptとは？ OpenAI Agents SDK TypeScript is a developer toolkit for building agentic AI applications with a small set of primitives. Agents combine language models with instructions and tools, while native TypeScript supports orchestration without requiring a separate workflow language. The SDK supports text tasks, agents that work in isolated filesystem workspaces, and realtime spoken interactions. It brings execution, delegation, validation, and tracing into one toolkit, with OpenAI model support included by default and an adapter for supported Vercel AI SDK models. ## OpenAI Agents SDK TypeScriptでできること ### Built-in agent loop Handles tool invocation, returns tool results to the language model, and continues execution until the task is complete. TypeScript functions can become tools with automatic schema generation and schema-based validation. ### Sandbox workspaces Provides isolated filesystems, shell commands, file editing, snapshots, and sandbox session state for tasks that need a persistent working environment. ### Realtime voice agents Supports low-latency spoken interactions with automatic interruption detection and context management. In browsers, RealtimeSession can connect microphone input and audio output through WebRTC. ### Delegation and tool connections Coordinates specialized agents through agents as tools and handoffs. Built-in MCP integration exposes server tools alongside local function tools. ### Validation and human involvement Runs input guardrails in parallel with agent execution and fails fast when checks do not pass. Human-in-the-loop mechanisms allow people to participate across agent runs. ### Tracing and session memory Built-in tracing visualizes, debugs, and monitors agent workflows. Sessions maintain working context within an agent loop, while tracing supports OpenAI evaluation, fine-tuning, and distillation tools. ## 利用シーン ### Repository inspection and editing Give a workspace assistant a Git repository to inspect, summarize, or modify using sandbox files and shell commands. Snapshots and session state support work that spans longer tasks. ### Spoken assistants in browser applications Build an assistant that accepts microphone input and responds with audio while using tools and preserving conversation history. Interruption handling supports back-and-forth spoken exchanges. ### Workflows with specialized agents Route parts of a larger task to agents with different instructions and tools, using handoffs or agent-as-tool calls to coordinate their contributions. ## よくある質問 ### How do I install the SDK? Run \`npm install @openai/agents zod\`. The SDK requires Zod v4. For the documented OpenAI text and sandbox examples, set the OPENAI\_API\_KEY environment variable. ### Which package should I start with? Use @openai/agents for most text, sandbox, and realtime applications. Choose @openai/agents-realtime when you only need the standalone realtime package. Lower-level packages support custom provider wiring and specific integrations. ### Can I use models from other providers? Yes, through the Vercel AI SDK adapter in @openai/agents-extensions. It supports AI SDK provider specification versions v2, v3, and v4. The adapter is in beta; the documentation recommends testing your chosen provider and using the default provider for OpenAI models. ### What limitations apply to the AI SDK adapter? It does not support deferred Responses tool loading, tool search, or Programmatic Tool Calling. PDF inputs can use base64 data or public URLs, subject to provider support; OpenAI file IDs are not supported through the adapter. ## クイック情報 - **Product type**: Agent application SDK - **Primary language**: TypeScript - **Default OpenAI APIs**: Responses API and Chat Completions API - **Main npm package**: @openai/agents - **Documentation**: https://openai.github.io/openai-agents-js