🇺🇸 USA · Claude Agent SDK
Status: 🟩 COMPLETE 🟦 LIVING Last updated: 2026-06-26 Plain-English tagline: Anthropic’s official software development kit for building production agents on top of Claude — the developer toolkit behind Claude Code itself.
Front-matter facts
| Field | Value |
|---|---|
| Vendor | Anthropic (San Francisco, USA) |
| Country / origin | 🇺🇸 USA |
| Recommended for Australian users? | ✅ Yes — open developer toolkit, accessible globally |
| Privacy summary | SDK uses the Claude API under the hood — same no-training-by-default API posture applies |
| Free tier | The SDK itself is free; you pay for the Claude API tokens it consumes |
| Paid tiers | API metered pay-per-token at standard Claude API rates |
| First released | 2025 — formally launched alongside the Claude Code Skills + Plugins ecosystem |
| Last reviewed | 2026-06-26 |
| Official site | https://docs.claude.com/en/docs/build-with-claude/agent-sdk |
What it is
The Claude Agent SDK is Anthropic’s official toolkit for building production AI agents on top of Claude. It provides:
- Pre-built agent primitives — message handling, tool calling, multi-turn conversation management
- MCP integration — easy connection to Model Context Protocol servers
- Skills support — your agents can use Claude Skills
- Sandboxing helpers — for agents that need isolated execution
- Streaming, error handling, retry logic — production-grade infrastructure
- Multi-model support — primarily Claude (Opus, Sonnet, Haiku, Fable) but extensible
- Available in Python and TypeScript / Node.js
Important framing: The Claude Agent SDK is the underlying toolkit that powers Claude Code itself. Claude Code is essentially “the most thoroughly-built agent product on the Agent SDK.” When you use Claude Code, you’re using a refined application of the SDK. When you build your own agent, you’re using the same building blocks.
This makes it analogous to:
- OpenAI Agents SDK — OpenAI’s equivalent for building agents on GPT models
- LangChain / LangGraph — third-party framework (more agnostic, more abstraction)
- CrewAI — third-party multi-agent orchestration
What you’d use it for
- Custom AI assistants for specific business workflows
- Internal agent tools at your organisation
- Embedded AI features inside larger products
- Multi-step workflows that require tool use, planning, retry logic
- Building MCP-integrated tools that work with both Claude Code and custom Claude clients
- Migrating from LangChain / similar frameworks when you want a more Anthropic-native posture
How to use it
Python
pip install anthropic-agent-sdkTypeScript / Node.js
npm install @anthropic-ai/agent-sdkQuick example (Python)
from anthropic_agent_sdk import Agent
agent = Agent(
model="claude-sonnet-4-6",
system="You are a helpful coding assistant.",
tools=[my_custom_tool, another_tool],
)
response = agent.run("Help me refactor this function")Real-world structure
- Define your agent’s tools (functions Claude can call)
- Configure system prompt + behaviour
- Optionally add MCP server connections
- Optionally add Skills
- Run the agent loop — SDK handles the multi-turn / tool-call / planning machinery
How it compares to alternatives
| Capability | Claude Agent SDK | OpenAI Agents SDK | LangChain / LangGraph | Pydantic AI |
|---|---|---|---|---|
| Native to a frontier lab | Yes (Anthropic) | Yes (OpenAI) | No (third-party) | No (third-party) |
| Multi-model support | Claude-focused | OpenAI-focused | Most agnostic | Multi-provider |
| MCP support | Native (deep) | Yes (newer) | Yes | Yes |
| Type safety | Python + TypeScript | Python + TypeScript | Python + JS | Best (Pydantic-backed) |
| Production-readiness | Yes (powers Claude Code) | Yes | Yes (with effort) | Yes |
| Learning curve | Moderate | Moderate | Steepest (many abstractions) | Gentle |
If you’re building for Claude specifically, the Agent SDK is the most direct path. If you want provider-agnostic code that could swap between Claude / GPT / Gemini, LangChain or LiteLLM are stronger choices. If you prefer strong typing in Python, Pydantic AI is excellent.
Privacy / data handling
- SDK uses the standard Claude API — no-training-by-default on inputs
- You’re responsible for handling user data in your custom agent — privacy posture is what you design
- For sensitive workloads, use AWS Bedrock or Google Cloud Vertex AI to route SDK API calls with regional data residency
Recent changes
- 2026: SDK expanded with richer Skills integration; better MCP server discovery
- 2025: Formal SDK launch (previously, building on Claude required more manual API code)
Gotchas
- The Agent SDK is the toolkit for building Claude Code clones — if you want to use Claude as an assistant rather than build a custom one, you don’t need the SDK
- Costs are metered API tokens — every agent step is API tokens; design tools to minimise wasted calls
- Tool call accuracy depends on tool descriptions — clear, specific tool descriptions matter as much as code quality
- Multi-turn loops can run away — implement turn limits and budgets
- Streaming requires careful client handling — server-sent events / chunked responses, not standard HTTP request/response
See also
- Claude Code deep dive 🟩 🟦
- Claude API overview 🟩 🟦
- Claude Skills 🟩 🟦
- Anthropic Computer Use 🟩 🟦
- Claude Cowork 🟩 🟦
- Claude Remote Desktop 🟩 🟦
- Subagents 🟩 🟦
- Tool use (concept) 🟩 — what’s underneath
- Agents (concept) 🟩
- MCP — Model Context Protocol 🟩 🟦
- OpenAI Agents SDK 🟥 — competing SDK
- LangGraph 🟥 — third-party framework
- Pydantic AI 🟥