Source Snapshot
- Origin: Anthropic Claude Agent SDK repositories, Claude Quickstarts, Anthropic Skills, and selected Claude Code ecosystem catalogs
- Type: Repository research note
- Author / org: Anthropic and community maintainers
- One-line takeaway: Study official SDK and quickstarts first, then use skills and subagent catalogs as patterns after security and permission review.
Garden Card
This note is a Quartz-ready learning map for Claude Agent SDK related GitHub repositories.
-
Core question: Which repositories should anchor SDK learning, prototyping, and enterprise agent design?
-
Operational value: It prevents jumping into community patterns before understanding official runtime boundaries.
-
Best connection: Agent SDK Core Concepts, SDK Demo Learning Map, How to Use Claude Cowork Effectively
1. Executive Summary
The official Python and TypeScript SDK repositories define the runtime contract: message streaming, options, tools, hooks, sessions, and client behavior. Demos and quickstarts make those primitives easier to inspect and extend.
Skills and subagent catalogs are useful after the baseline is understood. They should be treated as patterns to adapt, not production components to trust by default.
-
Main idea: Official SDK primitives come before ecosystem reuse.
-
Why now: Agent work is becoming reusable platform work.
-
Where it applies: Coding assistants, Feishu workflows, publishing agents, and manufacturing process assistants.
Decision Signal
Use official Anthropic repositories as the implementation baseline; use community catalogs as design inspiration only after checking permissions, code quality, and maintenance status.
2. Key Technical Terms
Use these terms when comparing SDK repositories and ecosystem assets.
-
SDK baseline: Official repository that defines the stable implementation surface.
-
Quickstart: Deployable example that shows application packaging and environment shape.
-
Skill: Reusable capability folder containing instructions, scripts, references, or assets.
-
Subagent: Specialized role with separate context, tools, and responsibility.
-
Pattern library: Collection of examples useful for design inspiration but not automatically trusted.
3. Core Notes
3.1 Problem
The Claude ecosystem contains official SDKs, demos, quickstarts, skills, and community catalogs. Without a learning order, it is easy to copy advanced patterns without understanding the runtime boundary.
-
Popularity does not equal production readiness.
-
Community assets may hide tool or data-access risk.
-
Official repositories reveal the real control surface.
3.2 Mechanism
Use a maturity ladder: SDK primitives, small demos, deployable quickstarts, reusable skills, then subagent role decomposition.
-
Start with read-only loops.
-
Add tools and hooks only after observing the message stream.
-
Review community assets before adoption.
3.3 Evidence
The official SDK repositories expose async query patterns, clients, MCP servers, hooks, option controls, error types, and examples.
-
Demos teach mechanics.
-
Quickstarts teach application shape.
-
Skills teach capability packaging.
3.4 Boundary
Do not treat repository examples as production-safe without review. Security, permissions, credentials, logging, licensing, and maintenance all need inspection.
-
Pin SDK versions before serious builds.
-
Audit custom tools and hidden instructions.
-
Separate learning experiments from production workflows.
4. Concept Map
Use wikilinks to connect this note into the broader Quartz graph.
- Related concept note: Agent SDK Core Concepts
- Related demo path: SDK Demo Learning Map
- Related operating guide: How to Use Claude Cowork Effectively
flowchart LR A["Official SDK"] --> B["Small Demos"] B --> C["Quickstarts"] C --> D["Skills"] D --> E["Subagents"] E --> F["Enterprise Agent Platform"] G["Community Catalogs"] --> E G --> H["Review Before Adoption"]
Diagram labels stay in English for rendering consistency and easier reuse across published pages.
5. My Take
The safest learning path is official-first, read-only-first, and governance-first. Reuse is valuable only after the runtime boundary is visible.
-
What changed my thinking: The best repositories show boundaries, not just features.
-
What I may do next: Build a read-only repository analyst before enabling write tools.
-
What still needs verification: Current SDK APIs, licenses, dependency health, and production security posture.
Reuse Path
Convert this note into a repository-selection checklist for Claude Agent SDK learning.