All Products
Search

This Product

    AgentCore:AgentCore SDK Overview

    Document Center

    AgentCore:AgentCore SDK Overview

    Last Updated:Sep 24, 2026

    When you build an agent application on AgentCore, the platform manages the models, MCP tools, Skills, memory, and credentials that the application uses. If you piece together a resource endpoint and an access key for each of them in your application code, the application stops working when the runtime changes, and credentials easily end up in source code or logs. AgentCore SDK lets you call these managed resources by name from your code, connect them to the agent framework you already use, and expose your agent as a service over the AG-UI or OpenAI Chat Completions protocol.

    This topic is the entry point to the SDK documentation. It helps you confirm what the SDK covers, choose a development language, prepare the resources you need, and select a serving protocol.

    Installation commands, code examples, and framework integrations are documented per language in the two user guides.

    Features

    AgentCore SDK provides the following features:

    Feature

    Description

    Models

    Call a cloud model by its model connection name, with both standard calls and streaming output. You can also connect your own model service.

    MCP

    List tools, call tools, and connect tools to your agent framework. Platform-managed MCP servers and your own services are both supported.

    Skill

    Load a Skill published on the platform, or load Skill files distributed with your application.

    Memory

    Add, retrieve, query, update, and delete memories, and connect them to the agent execution flow through a framework adapter.

    Credentials

    Retrieve a managed credential by name, or explicitly bind an MCP Header credential to an MCP server.

    Framework adapters

    Convert models, tools, and execution events into the corresponding objects and events that your agent framework recognizes.

    Serving protocols

    Use AgentCoreServer to serve AG-UI, OpenAI Chat Completions, or a custom protocol, or keep using your existing web service.

    Choose a development language

    Development language

    User guide

    Base runtime

    Python

    AgentCore Python SDK User Guide

    Python 3.10 or later

    JavaScript / TypeScript

    AgentCore Node.js SDK User Guide

    Node.js 20.3 or later

    Some frameworks require a later version of the base runtime, which the guide for your language specifies.

    Both guides cover how to call managed and custom resources, connect an agent framework, and deploy and call an agent service. Each guide also includes a complete LangChain agent service example, protocol invocation examples, and an FAQ.

    Framework and memory adapter support differs between the two languages. Each guide lists the frameworks and memory adapters that its language supports. If your application depends on a specific framework or memory adapter, compare these lists before you choose a language. Otherwise, read only the guide for the language you use. You do not need to cross-check the parameters or code of the other guide.

    Prerequisites

    1. Activate AgentCore and create the Workspace that your agent belongs to by following Manage Workspace.

    2. Prepare the resources your application needs and confirm that your agent has the required access permissions on them.

    3. Follow the user guide for your language to install the SDK and the framework dependencies you need, and then write and deploy your agent application. The alibabacloud-agentcore-sdk package itself does not deploy applications; the guide describes the deployment process.

      The following table lists the resources to prepare for each capability:

    Capability you need

    Resources to prepare

    Call a cloud model

    A model connection, plus a model configured in that connection. Manage Models describes how to create both.

    Use MCP

    An MCP server. Manage MCP Servers describes how to create one.

    Use Skill

    A Skill that is enabled and has a published version. Manage Skill describes how to create and publish it.

    Use memory

    A MemoryStore.

    Use managed credentials

    An API Key credential or an MCP Header credential, created as needed. Manage credentials describes how to create and authorize them.

    If your application only calls models, you do not need to create MCP servers, Skills, MemoryStores, or additional credentials.

    Choose a serving protocol

    Choose a serving option based on your requirements:

    Requirement

    Serving option

    Show the multi-step execution of your agent, message boundaries, tool calls, and tool results

    AG-UI

    Connect a client that uses the Chat Completions format

    OpenAI Chat Completions

    Serve a protocol other than AG-UI and OpenAI Chat Completions

    A custom protocol, served with AgentCoreServer

    Keep using the existing web service of your application

    Your existing web service

    AG-UI and OpenAI Chat Completions express different levels of detail. AG-UI keeps tool results as separate events, while OpenAI Chat Completions does not provide a separate response event for tool results. The "Call the Agent service" section of each language guide gives request examples, event descriptions, and limits.

    References