Anthropic offers several building blocks for creating agents with Claude. They can look confusing at first, but each has a clear role. Here is how they fit together.
Claude Agent SDK: the Claude Code engine in your app
The Claude Agent SDK is a Python and TypeScript library that runs the same agent loop that powers Claude Code, inside your own application. It comes with built-in tools for reading and editing files, running commands and searching the web, and lets you add your own tools, subagents and permission rules. It is the best choice for prototypes, internal tools and agents that work with your own files and systems.
import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions
async def main():
options = ClaudeAgentOptions(
system_prompt="You are an assistant for a small online shop.",
allowed_tools=["Read", "WebSearch"],
)
async for message in query(prompt="Summarise the questions in faq.md", options=options):
print(message)
asyncio.run(main())This example lets Claude read a local file and search the web, and nothing else. Limiting tools like this is the simplest and most important safety step. Check the official documentation for the current package version.
Claude Managed Agents: hosted by Anthropic
Claude Managed Agents, launched in April 2026, is a hosted API where Anthropic runs the agent loop, the secure sandbox and the session history for you. It suits long-running tasks, many users at once and integrations that need secure login tokens. A common path is to prototype with the Agent SDK and move to Managed Agents for production.
Agent Skills: reusable know-how
A skill is a folder with a SKILL.md file that teaches an agent how to do a specific task: the steps, rules and examples, plus optional scripts or templates. The agent loads a skill only when it is relevant, which keeps it focused. Agent Skills became an open standard in December 2025, and in October 2026 Anthropic donated it to the Agentic AI Foundation, so the same skill can work in more than 70 agent platforms.
---
name: weekly-sales-report
description: Builds the weekly sales report from orders.csv. Use when asked for the weekly report.
---
1. Read orders.csv and keep only the last 7 days.
2. Calculate total sales, number of orders and the top 5 products.
3. Write a short summary in English and Armenian.MCP: connecting Claude to your apps
The Model Context Protocol (MCP) is an open standard for connecting AI to apps and data, such as Google Drive, a CRM, a database or your website. Instead of writing a separate integration for every model, you build or install one MCP server, and any compatible agent can use it. The MCP specification released in July 2026 made servers simpler to scale and added interactive app interfaces and long-running tasks.
Which to choose
- Want to automate your own work without code: use Claude Cowork with skills and connectors.
- Building an internal tool or prototype: use the Claude Agent SDK.
- Launching an agent for many customers: use Claude Managed Agents.
- Need to connect any agent to your data: build or install an MCP server.
Frequently asked questions
Is the Claude Agent SDK free?
The library itself is free to use. You pay for Claude usage through the API or your plan, according to Anthropic’s current pricing.
Can I use Agent Skills with models other than Claude?
Yes. Agent Skills is an open standard, and many agent platforms support it. A well-written skill is just instructions and files, so it is easy to reuse.
Is MCP only for Claude?
No. MCP was created by Anthropic but is now an open standard under the Agentic AI Foundation, and it is supported by OpenAI, Google, GitHub, Apple and many others.
Read next
- How to Build an AI Agent: A Step-by-Step Guide for 2026
- Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5: What’s New and Which to Choose
- What Is an AI Agent? A Simple Explanation with Examples
Want an AI agent for your business? Tell us about your task. AI agents and automation
