Claude Agent SDK Katas

A port of the local-ai-agents-workshop katas from the Strands Agents SDK to the Claude Agent SDK — Anthropic's own SDK for building Claude Code-style agents in Python.

Why this exists

The original workshop teaches agents via Strands. The Claude Agent SDK is Anthropic's first-party alternative — it ships the same loop that powers Claude Code, with built-in tools (Read/Write/Bash/etc.), MCP support, and session management. These katas are the same exercises, ported one by one.

Status

Kata	Original (Strands, Python)	This repo — Python	This repo — TypeScript
02 — Agents Intro	✅	✅ ported	✅ ported
03 — Custom Tools	✅	✅ ported	✅ ported
04 — Local RAG	✅	not started	not started

SDK-only katas (no Strands counterpart)

These exercises show capabilities the Claude Agent SDK ships that Strands has no equivalent for. TypeScript only by request.

Kata	What it shows	Status
sdk-01 — Agent loop + PreToolUse hook	The SDK's built-in agent tools (Read/Glob/Grep/Bash/...) used with zero custom tools, observed and gated by a hook	✅ ported
sdk-02 — Subagents	Task tool + agents: {...} definitions, per-subagent tool/model scoping, observed via SubagentStop hook	✅ ported
sdk-03 — Lifecycle hooks	UserPromptSubmit + PostToolUse + Stop wired in one scenario — policy injection, output truncation, response-shape enforcement	✅ ported

Setup

Python

# Python 3.10+ required by claude-agent-sdk
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
export ANTHROPIC_API_KEY="sk-ant-..."
python kata-02-claude-sdk-intro/solution.py

TypeScript

cd kata-02-claude-sdk-intro-ts
npm install
export ANTHROPIC_API_KEY="sk-ant-..."
npm start

Both flavors require the Claude Code CLI (claude on PATH) — the SDK shells out to it.

Mapping from Strands

Strands concept	Claude Agent SDK equivalent
Agent(model=..., system_prompt=...)	ClaudeAgentOptions(model=..., system_prompt=...)
agent("hello") (one-shot)	query(prompt="hello", options=...)
agent("...") (multi-turn, same instance)	ClaudeSDKClient(options=...) session
@tool decorated Python function	@tool + in-process MCP server (create_sdk_mcp_server)
tools=[...] on the Agent	mcp_servers={...} + allowed_tools=[...] in options
Strands tool registry	MCP protocol (also works with external MCP servers)

The biggest mental shift: Claude Agent SDK is agent-loop-first. Its default toolset is Read/Write/Bash/Glob/Grep/etc. For purely conversational katas we explicitly pass allowed_tools=[] to keep the model from reaching for the filesystem.