Claude Agent SDK: What It Is, What You Can Build, and How It Fits the AI Agent Ecosystem

Last verified: March 2026. This guide is reviewed quarterly. Version-sensitive details are marked accordingly.

At a Glance: Claude Agent SDK

What it is	Anthropic's toolkit for building production-grade AI agents, built on the same infrastructure that powers Claude Code
Launched	September 29, 2025, alongside Claude Sonnet 4.5
Languages	Python (3.10+) and TypeScript / Node.js (18+)
Python install	pip install claude-agent-sdk (Claude Code CLI bundled; no separate install needed)
TypeScript install	npm install @anthropic-ai/claude-agent-sdk
Built-in tools	Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch, Skill
Model support	Claude models only (architectural constraint as of March 2026)
MCP integration	Native in-process server support
Directory listing	Claude Agent SDK on AI Agents Hub

Name change notice

The Claude Agent SDK was previously released as the Claude Code SDK. Anthropic renamed it in late 2025. According to the official migration guide, "The Claude Code SDK was originally designed for coding tasks, but it has evolved into a powerful framework for building all types of AI agents." The rename was not cosmetic: it accompanied real additions to the feature set, including subagents, lifecycle hooks, and the Skills system. If you are upgrading from the Claude Code SDK, the official migration guide covers the package name changes and settings behavior differences in detail.

What Is the Claude Agent SDK?

The Claude Agent SDK is Anthropic's open-source toolkit for building production-grade AI agents. Per Anthropic's documentation, the SDK lets you "build production AI agents with Claude Code as a library." That framing is deliberate: the SDK is built on the same infrastructure Anthropic's own team runs internally to power Claude Code.

Anthropic launched the SDK on September 29, 2025, alongside Claude Sonnet 4.5. The launch announcement explained what that internal foundation means in practice: "We've spent more than six months shipping updates to Claude Code, so we know what it takes to build and design AI agents. We've solved hard problems: how agents should manage memory across long-running tasks, how to handle permission systems that balance autonomy with user control, and how to coordinate subagents working toward a shared goal." Developers who adopt the SDK inherit those solved problems as a starting point.

The SDK is available in two languages:

• Python: pip install claude-agent-sdk (requires Python 3.10 or higher). As of the current release, the Claude Code CLI is bundled with the package. You do not need to install it separately, though you can specify a custom CLI path if preferred.
• TypeScript / JavaScript: npm install @anthropic-ai/claude-agent-sdk (requires Node.js 18 or higher).

Both packages are actively maintained. The npm package has published over 130 versions and accumulated over 600 dependents as of March 2026. The Python repository had approximately 4,800 GitHub stars as of March 2026.

What the SDK Enables

The Claude Agent SDK is broader than its "Code" origins suggest. According to Anthropic, developers are already building financial compliance agents, cybersecurity agents, and code debugging agents with the SDK. Anthropic also uses the SDK internally for GitHub issue triage and Slack automation workflows, which directly informed the SDK's design decisions around long-running tasks and permission systems.

Here are the concrete use case categories the SDK supports today:

• Coding and development agents. The original use case. Agents that read codebases, write files, run tests, and automate pull request workflows.
• Business workflow agents. Legal document review, financial compliance checks, customer support escalations, and multi-step approval workflows.
• DevOps and infrastructure automation. GitHub issue triage, CI/CD pipeline agents, automated incident response, and Slack bot integrations.
• Research and analysis agents. Agents with web access via the built-in WebSearch and WebFetch tools, capable of gathering, synthesizing, and reporting on external information.

The built-in toolset and native MCP integration layer make these use cases viable without requiring you to wire up file access, web search, or subprocess management from scratch.

Core Capabilities and Built-in Tools

Two usage modes

The SDK provides two primary ways to interact with Claude:

query() is the simpler entry point for single-turn or lightweight use cases. It works well when you need a straightforward model response without managing session state.

ClaudeSDKClient is the mode to use for full tool access and multi-turn sessions. It maintains session state across turns and supports the full built-in toolset. If your agent needs file operations, web access, code execution, or multi-step reasoning, use ClaudeSDKClient.

A minimal ClaudeSDKClient usage pattern in Python:

import asyncio
from claude_agent_sdk import ClaudeSDKClient

async def main():
    async with ClaudeSDKClient() as client:
        async for message in client.run("Summarize the files in the current directory"):
            print(message)

asyncio.run(main())

For production use, configure allowed_tools, permission_mode, and optionally setting_sources based on your agent's requirements.

Built-in tools

The core built-in tools as of March 2026:

Read | Write | Edit | Bash | Glob | Grep | WebSearch | WebFetch | Skill

You do not need to implement file operations, shell execution, or web access yourself. These tools are available via the allowed_tools parameter. Note that this is the core list; Anthropic adds tools in SDK updates, so the full set may be larger in current releases.

In-process MCP server support

The SDK supports running MCP (Model Context Protocol) servers within the same Python or TypeScript application, rather than as separate server processes. According to the GitHub README, this removes the need to manage subprocesses and reduces latency from inter-process communication for tool calls. For teams already working with MCP server integrations, this means simpler deployment and faster execution compared to architectures that require separate server processes.

Settings isolation and Skills

By default, the SDK does not load CLAUDE.md files, Skills, or project-level settings. To enable these, set setting_sources=['project'] in Python (or settingSources: ['project'] in TypeScript) in your agent options.

The Skills feature allows you to define reusable agent capabilities as SKILL.md files in .claude/skills/. Skills are automatically discovered at startup and are model-invoked: Claude selects when to use them based on the task at hand. This requires explicit configuration via setting_sources.

Claude Agent SDK vs. Other Frameworks

If you are evaluating the Claude Agent SDK alongside other frameworks, the most useful comparison is at the architectural level, not the feature checklist level. Frameworks differ in how they handle model dependencies, orchestration, and tool integration in ways that matter significantly once you reach production.

The following comparison draws from framework analysis published in early 2026, including reviews from letsdatascience.com and gurusup.com. These are editorial assessments, not benchmarks or Anthropic-official characterizations.

Framework comparison guides published in early 2026 characterize the Claude Agent SDK as the strongest choice for MCP-native agent development, noting its in-process server model and lifecycle hooks as distinguishing features.

The SDK is not a universal framework choice. Two things are worth stating directly:

The model constraint is real. As of March 2026, the SDK is designed for Claude models specifically. If model-provider flexibility is a current requirement for your project, this is an architectural constraint to evaluate before committing. Teams building on LangGraph or crewAI retain the option to swap underlying models without rebuilding their orchestration layer. LangGraph in particular handles complex stateful workflows with more mature multi-agent coordination than the SDK offers today.

The Claude commitment has benefits. If your team is already on Anthropic's model family, this constraint becomes a feature: the SDK is optimized for Claude-specific behaviors including extended thinking, tool use, and the computer-use capabilities. You are not writing around abstraction layers built for multiple models.

If you are also evaluating AutoGen, it is worth noting that AutoGen's conversation-first orchestration model differs significantly from the tool-use-chain approach of the Claude Agent SDK. The right choice depends more on your orchestration requirements than on any single feature comparison.

Who Should Use the Claude Agent SDK?

The SDK is not the right tool for every project. Here is a neutral framing to help you decide.

You are building on Claude Code. If you are already working within the Claude Code environment and want to extend its capabilities programmatically, the SDK is the natural next step. It uses the same underlying infrastructure, the same tool configuration patterns, and the same Skills system. The transition from Claude Code to SDK-based automation is lower friction than switching to a general-purpose framework.

You are building MCP-native agents. If your agent architecture relies heavily on MCP servers for tool integration and you want in-process execution without managing separate server processes, this is the framework with the deepest native MCP support as of early 2026. Framework comparison guides consistently highlight this as the SDK's architectural differentiator.

You are evaluating before committing. If you are prototyping and have not yet chosen a model provider, consider the model-dependency constraint carefully before building deeply on this SDK. This is not a reason to avoid the SDK, but it is a design decision that is expensive to reverse once an agent system is in production. If you are already productive in LangGraph or crewAI and your project does not require deep MCP integration or Claude-specific optimization, there may not be a compelling reason to migrate today.

If you are still evaluating which framework or tool fits your stack, the AI Agents Hub directory listing for Claude Agent SDK includes hands-on assessment notes, GitHub stats, and related framework listings. Explore the Claude Agent SDK listing on AI Agents Hub.

The Claude Agent SDK Ecosystem

The SDK has moved well beyond a standalone open-source release. A set of official integrations, testing tools, and enterprise connections has formed around it since the September 2025 launch.

Official resources

Testing and evaluation

Promptfoo provides official provider support for the Claude Agent SDK, including support for testing Skills. It is a practical option for teams running systematic evaluations before deploying agents to production. In Promptfoo's eval mode, the SDK runs in a temporary directory by default. See the Promptfoo provider docs for configuration details.

Observability

LangSmith (LangChain's observability product) has an official Claude Agent SDK integration via langsmith.integrations.claude_agent_sdk. The configure_claude_agent_sdk() function enables full trace visibility across agent runs. This is useful for debugging long-running agents where tool call sequences need to be inspected across turns. See the LangSmith tracing docs for setup instructions.

Enterprise integrations

Microsoft Agent Framework has published an official Python package (agent-framework-claude) for integrating Claude agents into workflows alongside agents built with other providers. This means the Claude Agent SDK does not need to be an isolated implementation choice. Teams can mix Claude agents with agents from other frameworks within a single coordinated workflow. See the Microsoft Developer Blog post for an integration walkthrough.

If you want to be notified when we publish new agentic framework guides and SDK deep dives, sign up for the AI Agents Hub newsletter below. We cover new framework reviews, comparison breakdowns, and directory additions as they are published.

Explore the Claude Agent SDK listing and related agentic frameworks in the AI Agents Hub directory. The listing includes GitHub stats, capability notes, and links to related tools in the agentic frameworks category.