AWS Agent Toolkit — Agent Plugin [ experimental ]

A portable Agent Plugins package that bundles the AWS MCP Server and a set of curated AWS skills into one installable plugin, built from the Agent Toolkit for AWS (Apache-2.0).

It conforms to the Agent Plugins 1.0.0 specification, so any compatible client can discover and load it consistently.

What's inside

aws-agent-toolkit-plugin/
├── plugin.json                 # Agent Plugins manifest (identity + metadata)
├── mcp.json                    # AWS MCP Server configuration (stdio via mcp-proxy-for-aws)
├── README.md
└── skills/
    ├── working-with-aws/       # Baseline guardrails: MCP-first, IaC-first, verify vs docs, Well-Architected
    ├── aws-mcp-server/         # How to drive the AWS MCP Server (tools, auth, pagination, IAM condition keys)
    ├── s3-secure-storage/      # Step-by-step: secure, versioned S3 bucket + 90-day Glacier lifecycle (CDK + CFN)
    ├── serverless-api/         # Step-by-step: Lambda + API Gateway serverless API (SAM + CDK)
    └── choose-a-database/      # Decision guide: DynamoDB vs Aurora vs RDS vs DocumentDB vs Neptune, etc.

The two components map directly to the Agent Toolkit's pillars:

AWS MCP Server (mcp.json) — secure, IAM-governed access to 300+ AWS services and 15,000+ API actions, sandboxed Python execution, and live AWS documentation, through one authenticated endpoint with CloudWatch metrics and CloudTrail logging.
Agent Skills (skills/) — curated, load-on-demand instructions covering the three archetypes the toolkit describes: guardrail/rules, step-by-step procedures, and service decision guides. The working-with-aws skill folds in the toolkit's recommended rules file guidance.

The MCP server

mcp.json registers the AWS MCP Server via its stdio proxy:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json",
  "mcpServers": {
    "aws-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "mcp-proxy-for-aws@1.6.3",
        "https://aws-mcp.us-east-1.api.aws/mcp",
        "--metadata",
        "AWS_REGION=us-west-2"
      ]
    }
  }
}

Change AWS_REGION to your working region.
Pin the proxy version (e.g. @1.6.3) for reproducibility; the upstream repo currently ships @1.6.4. Bump deliberately after checking PyPI.

Schema note (portable vs client-native)

The AWS product page and the upstream aws-core plugin write the server in the Claude/Cursor client-native form ("transport": "stdio", plus a "timeout" field, and no $schema). The Agent Plugins mcp.schema.json is a closed schema (additionalProperties: false) that instead requires a "type" field and forbids unknown keys. This package uses the portable form: $schema added, transport → type, and the non-portable timeout dropped.

Prerequisites

uv installed (runs the MCP proxy via uvx).
AWS credentials — optional, and only for API calls. Documentation search and skill retrieval work unauthenticated. Configure with aws configure or aws sso login and reference profiles by name; the plugin never stores credentials.

Install

Point your client at this directory (it is a self-contained Agent Plugin). For example, in an MCP-capable agent, merge mcp.json into your MCP configuration, and load the skills from skills/. See your client's Agent Plugins loading docs for the exact install step.

Try it

Ask your agent:

"Create an S3 bucket with versioning enabled and a lifecycle policy that transitions objects to Glacier after 90 days."

It should load the s3-secure-storage skill and return reviewable CDK or CloudFormation with secure defaults.

Provenance & license

Derived from the open-source Agent Toolkit for AWS and repackaged into the portable Agent Plugins format. The skill content here is original, written to reflect the toolkit's approach. Licensed under Apache-2.0, matching the upstream project. "AWS", "Amazon Web Services", and related marks belong to Amazon.