AppNewofficial publisher

Claude Agent SDK

App : A product you use directly

The engine behind Claude Code made available to developers: files, terminal, web search, subagents, skills, and fine-grained permissions in your own agents.

The most complete option for agents working on files and code. Limited to Claude models, via the Anthropic API, Amazon Bedrock, or Google Vertex AI.

Official site GitHub · 8,293 ★ Docs

Overall score

8.1 /10

Security: one caveat, see the card

Utility

9.0

Reliability

10.0

Security

One caveat

Ease

5 min · Technical

Security reflects where data goes, requested permissions and incidents on record at the verification date. It is not a guarantee of zero risk: check the deployment mode and the connections you enable.

What it does, concretely

Building an AI agent capable of reading files, running commands, and browsing the web requires assembling every piece yourself: agent loop, permission management, sessions, subagents. The Claude Agent SDK provides exactly that engine, the one powering Claude Code, as a Python (and TypeScript) library you embed in your own applications.

In practice, you install the package via pip, call the query() function or the interactive ClaudeSDKClient, and configure allowed tools (read, write, terminal, web search, MCP). You can define custom tools in plain Python, hooks that run before or after each tool call, and spawn specialized subagents for focused subtasks. The Claude Code CLI is bundled in the package, so no separate installation is needed.

This SDK targets Python or TypeScript developers who want to build production agents that work on code or files, without rewriting the agent loop from scratch. Teams already using the Anthropic API, Amazon Bedrock, or Google Vertex AI will find direct integration.

It is not suited for those looking for a graphical interface, a no-code approach, or who want to use models other than Claude. People who prefer to call the Claude API directly and manage the tool loop themselves should look at Anthropic's Client SDK instead.

What it offers

·Built-in tools to read, write, edit files, run shell commands, and perform web searches.
·Granular permission system via allowed_tools and disallowed_tools to control which tools run automatically or require approval.
·Python hooks executed by the application (not by Claude) at specific points in the agent loop, enabling deterministic control over actions.
·Custom tools defined as Python functions and exposed to Claude via in-process MCP servers, with no separate subprocess required.
·Specialized subagents to delegate focused subtasks within a multi-agent workflow.
·Session management with resume and fork capabilities to maintain context across multiple exchanges.
·Compatibility with external MCP servers (stdio) alongside internal in-process MCP servers.
·Support for skills, commands, and memory loaded from the project's .claude/ folders and ~/.claude/, just like Claude Code.

Our tips

Use disallowed_tools to explicitly block dangerous tools: allowed_tools does not remove a tool from the available set, it only pre-approves it.
Set snapshot to false in the system_prompt configuration so the prompt is rebuilt on every request during development, rather than being locked in after the first session.
Prefer in-process MCP servers (create_sdk_mcp_server) over external servers for your custom tools: less overhead, simpler debugging, single-process deployment.
Use PreToolUse hooks to intercept and block specific shell commands before they run, by returning a permissionDecision of deny with an explanatory message.
Specify the working directory via cwd in ClaudeAgentOptions to confine the agent to the relevant project folder and avoid unintended changes elsewhere.

Watch out

The SDK is limited to Claude models: no compatibility with other model providers, creating a direct dependency on Anthropic, Amazon Bedrock, or Google Vertex AI.
The system prompt is locked in on the first request of a session and reused as-is on subsequent requests, unless snapshot is explicitly set to false (requires Claude Code CLI 2.1.257 or later).

In practice

Price: Freemium
Commercial use: Commercial use: yes
Available on: cli
FR interface: No

Security and data

Rating: Security: one caveat, see the card
Your data: Local or cloud depending on use
Maintenance: Active
Last release: 15 October 2026

Links and repository checked on 18 October 2026 · added on 27 September 2026

How to install Claude Agent SDK

1. Make sure Python 3.10 or later is installed on your machine. 2. Install the SDK with pip install claude-agent-sdk (the Claude Code CLI is bundled in the package). 3. Set your Anthropic API key as an environment variable (ANTHROPIC_API_KEY). 4. Define allowed tools via allowed_tools and explicitly block unwanted ones with disallowed_tools before any deployment. 5. Test with a minimal script using query() on an isolated working directory before granting write or execution permissions.

pip install claude-agent-sdk

Official documentation ↗

In the same category

Whole category Automation & agents →