aws-skills — Summary

aws-skills is an Anthropic-format Claude Code plugin marketplace for AWS development, shipping 5 plugins (aws-common, aws-cdk, aws-cost-ops, serverless-eda, aws-agentic-ai) with 5 skills and bundled MCP server configurations covering CDK, serverless architecture, cost optimization, and Bedrock AgentCore. Each skill uses context: fork — it runs in a forked context to avoid polluting the main session. The aws-cdk-development skill has an inline PreToolUse hook that validates AWS credentials before every cdk deploy invocation (aws sts get-caller-identity). Skills specify allowed-tools: lists to constrain which MCP tools and Bash commands the agent may use. The aws-agentic-ai plugin provides comprehensive Bedrock AgentCore documentation covering Gateway, Runtime, Memory, Identity, Code Interpreter, Browser, and Observability services. With 294 stars and last commit April 2026.

Differs from seeds: No seed covers cloud infrastructure or AWS. Closest to spec-kit (skills with per-skill hooks + MCP integration), but the context: fork pattern and allowed-tools: constraint lists are AWS-specific features. The aws-common plugin's MCP-first instruction ("always verify AWS facts using MCP tools before answering") is the only example in the batch of mandatory MCP-before-response validation.

aws-skills — Overview

Origin

Created by zxkane (GitHub), an AWS practitioner. Current focus on CDK, serverless/event-driven architecture, cost operations, and the emerging Bedrock AgentCore platform. Licensed MIT.

Philosophy

From the README: "Claude Code plugins for AWS development with specialized knowledge and MCP server integrations, including CDK, serverless architecture, cost optimization, and Bedrock AgentCore for AI agent deployment."

The aws-cdk-development skill makes the philosophy explicit:

"Always verify AWS facts using MCP tools (mcp__aws-mcp__* or mcp__*awsdocs*__*) before answering."

This is a mandatory MCP-first pattern — the skill explicitly instructs the agent to check live AWS documentation before responding, preventing outdated advice. The aws-common dependency is automatically loaded by all other plugins, ensuring the MCP configuration is always available.

MCP-First Architecture

Unlike most frameworks that use MCP as optional tooling, aws-skills makes MCP verification mandatory before answering AWS questions. The aws-mcp-setup skill in the aws-common plugin provides step-by-step guidance for configuring MCP servers if they are unavailable.

Plugin Dependency Graph

aws-cdk → aws-common (auto-loaded)
aws-cost-ops → aws-common (auto-loaded)
serverless-eda → aws-common (auto-loaded)
aws-agentic-ai → aws-common (auto-loaded)

aws-skills — Architecture

Distribution

Anthropic-format plugin marketplace.

Install

# Register marketplace
/plugin marketplace add zxkane/aws-skills

# Install plugins
/plugin install aws-common@aws-skills     # dependency, install first
/plugin install aws-cdk@aws-skills
/plugin install aws-cost-ops@aws-skills
/plugin install serverless-eda@aws-skills
/plugin install aws-agentic-ai@aws-skills

# Alternative: install individual skills via npx
npx skills add https://github.com/zxkane/aws-skills --skill aws-cdk-development

Directory Tree

.
├── .claude-plugin/marketplace.json       # Marketplace catalog
├── .claude/skills/                       # Flat skill copies (alternate install)
│   ├── aws-agentic-ai/
│   ├── aws-cdk-development/
│   ├── aws-cost-operations/
│   ├── aws-mcp-setup/
│   └── aws-serverless-eda/
├── plugins/
│   ├── aws-agentic-ai/
│   │   └── skills/aws-agentic-ai/
│   │       ├── SKILL.md
│   │       ├── cross-service/
│   │       ├── references/
│   │       ├── scripts/               # Dockerfile, Python templates
│   │       └── services/              # browser, code-interpreter, gateway, identity, memory, registry
│   ├── aws-cdk/
│   │   └── skills/aws-cdk-development/SKILL.md
│   ├── aws-common/
│   │   └── skills/aws-mcp-setup/SKILL.md
│   ├── aws-cost-ops/
│   │   └── skills/aws-cost-operations/SKILL.md
│   └── serverless-eda/
│       └── skills/aws-serverless-eda/SKILL.md
├── docs/
├── CLAUDE.md
└── README.md

Required Runtime

• Claude Code
• AWS CLI (aws sts get-caller-identity)
• AWS MCP servers (various, configured via aws-mcp-setup)
• Node.js + CDK CLI (cdk deploy, for aws-cdk plugin)

aws-skills — Components

Skills (5)

Skill Features (aws-cdk-development)

context: fork          # Runs in forked context
skills:
  - aws-mcp-setup      # Auto-loads dependency
allowed-tools:
  - mcp__cdk__*
  - mcp__aws-mcp__*
  - mcp__awsdocs__*
  - Bash(cdk *)
  - Bash(npm *)
  - Bash(aws cloudformation *)
  - Bash(aws sts get-caller-identity)
hooks:
  PreToolUse:
    - matcher: Bash(cdk deploy*)
      command: aws sts get-caller-identity --query Account --output text
      once: true

Scripts (aws-agentic-ai plugin)

MCP Servers Referenced

• AWS CDK MCP (stdio)
• AWS Pricing MCP
• AWS Cost Explorer MCP
• Amazon CloudWatch MCP
• AWS Documentation MCP (mcp__aws-mcp__*, mcp__awsdocs__*)

Hooks (1 inline)

aws-cdk-development SKILL.md frontmatter:

• PreToolUse on Bash(cdk deploy*) — runs aws sts get-caller-identity once before first CDK deploy

aws-skills — Uniqueness

Differs from Seeds

No seed covers AWS or cloud infrastructure. The context: fork + allowed-tools: pattern is not present in any seed — it creates a scoped, sandboxed sub-context for CDK operations that prevents tool bleed-out. The inline skill-level hooks: declaration in SKILL.md frontmatter (rather than a separate hooks.json file) is a more ergonomic pattern than spec-kit's separate hook files. The mandatory MCP-before-answering instruction ("Always verify AWS facts using MCP tools before answering") is the only example in the corpus of making MCP verification a prerequisite for response generation rather than an optional enhancement.

Positioning

The only actively maintained AWS-focused Claude Code plugin. The aws-agentic-ai plugin is uniquely forward-looking — it covers Bedrock AgentCore, a platform that enables deploying and scaling AI agents on AWS, including protocols (A2A, AG-UI, MCP). This makes aws-skills simultaneously a CDK/serverless harness and a guide for building production AI agent infrastructure.

Observable Failure Modes

1. MCP server availability: Skills depend on AWS MCP servers that must be independently installed and configured; aws-mcp-setup provides guidance but does not auto-install.
2. Region-specific behavior: CDK patterns and service availability differ by AWS region; skills do not account for this.
3. Bedrock AgentCore private preview: The aws-agentic-ai plugin covers a platform that may have changed since the skills were written.

Explicit Antipatterns

• Explicit resource naming when optional (CDK anti-pattern — causes deployment conflicts)
• Not verifying AWS facts via MCP before answering (mandatory MCP check)
• Running cdk deploy without valid credentials