Claude Agent SDK: reuse the engine behind Claude Code

From a terminal tool to a library

It all starts with Claude Code, Anthropic’s coding assistant that works in the terminal. Its internals were first offered to developers as the Claude Code SDK, then renamed Claude Agent SDK in September 2025 to signal that it isn’t only for coding. The idea: instead of writing your own AI agent loop and tools from scratch, you reuse Claude Code’s.

Out of the box, the agent can read, write and edit files, run shell commands and search the web. On top of that you get specialized subagents, hooks (your own code triggered at key moments), a permission system that decides which actions need your approval, resumable sessions, and connections to MCP servers to plug in databases, browsers or APIs.

What it’s good for

The SDK shines when an agent has to act on a real environment: fixing bugs, writing and running tests (agentic coding), sorting documents, producing reports, or automating a step in a CI pipeline. Anthropic publishes sample agents on GitHub, such as an email assistant and a research agent.

Getting started

You need Node.js 18+ or Python 3.10+ and an Anthropic account. Install claude-agent-sdk (pip) or @anthropic-ai/claude-agent-sdk (npm), set your API key, then call the query function with a prompt and the list of allowed tools. The official quickstart walks you through an agent that finds and fixes bugs in a file.

Strengths and limits

Strengths: a heavily equipped agent from the first line of code, the same machinery as Claude Code, and fine-grained control over permissions and context.

Limits: it only works with Claude models; its use is subject to Anthropic’s Commercial Terms; and every run consumes billed tokens, which can add up on long tasks. An agent that can run commands needs tight boundaries: sandboxing and our can I trust an agent with this? tool help you scope the risks. If you’d rather have Anthropic host the agent than run it on your own servers, the company offers a separate product called Managed Agents.

Alternatives

The OpenAI Agents SDK and Google ADK play a similar role for other vendors, while smolagents and Pydantic AI are model-agnostic.