How Makers Build
Check my stackPlan my stack
Menu

Claude Agent SDK

Anthropic's SDK for building agents on the same harness as Claude Code — tool use, file and shell access, subagents and context management — in Python and TypeScript.

Visit code.claude.com ↗Alternatives
Works with AI agents:llms.txt
Ask your AI about this, with this page as the source:ChatGPT ↗Claude ↗Perplexity ↗

The Claude Agent SDK is a library that puts the agent behind Claude Code inside your own program. Instead of writing the tool-calling loop yourself, you call one function and get an agent that reads and edits files, runs shell commands, searches the web and decides when it is done.

The model to understand: when your code calls query(), the SDK starts a claude CLI process and talks to it over stdio. That process owns a shell, a working directory and session transcripts on local disk, so each running session is a long-lived process with local state, not a stateless API call. The SDK ships for Python and TypeScript; other languages can run the CLI as a subprocess with JSON output.

What a small team gets: built-in file, shell and web tools; your own functions exposed as tools through an in-process MCP server, plus external MCP servers; permission rules for which tools run without approval; hooks at points in the agent lifecycle; subagents; and sessions you can resume or fork. It calls the Claude API with an API key, or Claude through Amazon Bedrock, Google Cloud or Microsoft Foundry. You host it yourself, usually in sandboxed containers.

The main limit is operational. Anthropic suggests about 1 GiB of RAM, 5 GiB of disk and one CPU per agent as a starting point, and transcripts and working files disappear with the container unless you persist them. For a few tool calls per request, the plain Claude client SDK is lighter.

Where it fits

AI SDK & Agent Framework · Agent harnessAgents that read and edit files, run shell commands and manage long context, built on the same harness as Claude Code.Trade-off: Built for Claude models, and it needs a filesystem and process to run in, which serverless functions don't give you.

Who uses it

No maker's live product on record yet, and 58 open-source projects that declare it in their code.

Declared in code 58How evidence is collected →

We haven't found a maker's live product that uses Claude Agent SDK yet — only the open-source projects below, which declare it in their code.

Aix-DBAix-DB 基于 LangChain/LangGraph 框架，结合 MCP Skills 多智能体协作架构，实现自然语言到数据洞察的端到端转换。AI SDK & Agent FrameworkIn its code · source ↗

Open source: a project that declares Claude Agent SDK as a dependency in its public code — verifiable, but not necessarily a live product.

Loved and watch-outs

Themes that recur in makers' words and Hacker News comments, each linked to what it summarises.

Most loved
  • Exposes nearly everything Claude Code has, including subagents, plan mode, background tasks, plugins, shell and MCP tools, as a library. HNHN 2HN 3HN 4
  • A headless Claude Code makes custom desktop and mobile front ends far easier to maintain than scraping terminal output. HNHN 2HN 3
  • Teams stand up internal agents quickly, such as Slack helpers, grading tools and agent factories. HNHN 2HN 3HN 4
  • Hooks and permission controls are unusually granular for an agent SDK. HN
Watch-outs
  • Rules on using it with a Claude subscription are confusing and keep changing, including an announced billing change that was later paused. HNHN 2HN 3HN 4
  • It feels heavy and hard to customize, essentially a thin, closed wrapper around the CLI with incomplete docs. HNHN 2
  • Automations, scheduling and security boundaries need separate infrastructure you build yourself. HN
  • It ties you to Claude models, so some developers build vendor-agnostic replacements. HNHN 2

Reliability and open issues

Most wanted on GitHubOpen on 2026-10-04 in anthropics/claude-agent-sdk-python, with activity in the last year — issues and feature requests by 👍.

Alternatives to Claude Agent SDK

All alternatives by situation →
LangChainThe largest set of ready-made integrations for models, tools and data sources, in Python and JavaScript.Open sourceFree tier65 makers · AI SDK & agent framework
LlamaIndexAgents that mostly retrieve and reason over your own documents.Open sourceFree tier16 makers · AI SDK & agent framework
MastraTypeScript and Next.js apps that want agents, workflows, memory and RAG without a Python service.Open sourceFree tier24 makers · AI SDK & agent framework
OpenAI Agents SDKA small agent loop with tools, handoffs, guardrails and tracing built in, from the OpenAI team.Open sourceFree tier1 makers · AI SDK & agent framework
Pydantic AIPython agents whose outputs must validate against typed schemas, across any model provider.Open sourceFree tier
AI SDKTypeScript agents with tool calling, multi-step loops and streaming to the UI, on any model provider.Open sourceFree tier4 makers · AI SDK & agent framework
LangGraphStateful agents built as an explicit graph of steps, with checkpoints, retries, streaming and pauses for human approval, in Python or JavaScript.Open sourceFree tier1 makers · AI SDK & agent framework
CrewAIPython projects that split work across several role-based agents, like researcher, writer and reviewer.Open sourceFree tier3 makers · AI SDK & agent framework

Questions makers ask about Claude Agent SDK

Can I run it in a serverless function?

It is built for a container rather than a stateless function. Each session is a long-lived claude subprocess with its own shell, working directory and transcripts on local disk, and Anthropic's hosting guide sizes containers at roughly 1 GiB RAM, 5 GiB disk and one CPU per agent. source ↗

Can I use it through Amazon Bedrock or Google Cloud instead of the Anthropic API?

Yes. Set an environment variable for Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry and configure that cloud's credentials; otherwise it uses an Anthropic API key. source ↗

Can my users sign in with their Claude.ai subscription?

No, not without prior approval from Anthropic. Third-party products built on the SDK are expected to authenticate with API keys, not claude.ai login or its rate limits. source ↗

Which languages are supported?

Python (3.10+) and TypeScript (Node.js 18+). Both packages bundle the native Claude Code binary; from other languages you can drive the same loop by running the CLI as a subprocess with JSON output. source ↗

How do I give the agent my own tools?

Define functions with a name, description and input schema, wrap them in the SDK's in-process MCP server, and pass it in the query options. You can also connect external MCP servers and restrict which built-in tools the agent sees. source ↗

Do sessions survive a container restart?

Not by default; transcripts, memory files and working-directory files live on the container's disk. A SessionStore adapter can persist transcripts across hosts, but other files need a mounted volume or your own sync. source ↗

Is it safe to serve several customers from one container?

Only with care. By default the SDK loads settings and memory files from the filesystem, so Anthropic's guide recommends a separate working directory and config directory per tenant and disabling auto memory, or one sandboxed container per agent. source ↗

Is Claude Agent SDK free?

No permanent free tier; paid use starts at Pay per token. source ↗

Is Claude Agent SDK open source or self-hostable?

Open source, and you can self-host it. source ↗

Can AI coding agents work with Claude Agent SDK?

It serves an llms.txt docs index.

Who uses Claude Agent SDK?

No maker's live product we track yet; 58 open-source projects declare it in their code. source ↗