How is the Agent Toolkit different from awslabs/mcp?

The Agent Toolkit is the umbrella; awslabs/mcp is one component of it. Specifically: the AWS MCP Server (in aws/agent-toolkit-for-aws repo) is the new GA managed MCP server with enterprise governance (CloudWatch metrics, IAM context keys, CloudTrail audit). awslabs/mcp continues to host specialized MCP servers for individual AWS services (Bedrock Knowledge Bases, CDK, ECS, Lambda, DynamoDB, RDS) and is still actively maintained alongside the new toolkit. Use the AWS MCP Server (Agent Toolkit) for general AWS service access and governance; pull in awslabs/mcp servers for deeper service-specific coverage.

Which agents support the AWS Agent Toolkit?

Per AWS's launch documentation, the toolkit supports any MCP-compatible agent. AWS specifically lists Claude Code, Kiro, Kiro CLI, Cursor, Codex CLI, Cline, and Windsurf as confirmed working. In practice, every major coding and chat agent in May 2026 supports MCP — including ChatGPT Atlas, JetBrains Air, OpenClaw, Continue, Aider (via plugin), and Roo Code. The Agent Toolkit's value is that AWS does the integration testing for you and ships skills + plugins that pre-wire common AWS workflows, rather than you assembling MCP servers manually.

How do I install the AWS Agent Toolkit?

Install paths vary by agent. For Claude Code: 'claude mcp add aws-mcp-server' (or via the AWS-published plugin). For Cursor: install the AWS MCP Server plugin from the Cursor MCP marketplace or configure manually in mcp.json. For Codex CLI: configure the AWS MCP Server endpoint in Codex config. For Kiro: pre-installed natively (Kiro is AWS's own agentic IDE). The agent-toolkit-for-aws GitHub repo has full install commands for each supported client. Once installed, the agent gets AWS service tool calls (S3, EC2, IAM, Lambda, etc.) with CloudTrail-audited execution out of the box.

Quick Answer

What is the AWS Agent Toolkit? MCP, Skills, Plugins (May 2026)

Q: What is the AWS Agent Toolkit?

The Agent Toolkit for AWS is a bundle of official, AWS-supported MCP servers, skills, and plugins that AI coding agents (Claude Code, Cursor, Codex, Kiro, Cline, Windsurf) can use to build on AWS. It launched on May 6, 2026 alongside the AWS MCP Server going generally available. The toolkit's three pillars: (1) AWS MCP Server — the GA managed MCP server for AWS service access; (2) skills — declarative agent capabilities specifically for AWS workflows; (3) plugins — drop-in agent extensions for popular agents. The toolkit is the official AWS answer to 'how do AI coding agents safely build on AWS?'.

Published: May 7, 2026

What is the AWS Agent Toolkit? MCP, Skills, Plugins (May 2026)

On May 6, 2026, AWS launched the Agent Toolkit for AWS — a bundle of official, AWS-supported MCP servers, skills, and plugins that lets AI coding agents (Claude Code, Cursor, Codex, Kiro, Cline, Windsurf) build on AWS with enterprise governance. It’s the umbrella that includes the new AWS MCP Server going GA, the existing awslabs/mcp specialized servers, and a layer of skills and plugins for popular agent clients.

Last verified: May 7, 2026

The three pillars

The Agent Toolkit has three main components:

1. AWS MCP Server (GA May 6, 2026)

The flagship: a managed Model Context Protocol server that exposes AWS service APIs as MCP tool calls.

Hosted by AWS — no self-hosting required.
CloudWatch metrics under the new AWS-MCP namespace.
IAM context keys for agent-specific policies.
CloudTrail audit of every call.
Approval gates for resource-modifying actions.
Read-only mode as a default safety setting.

This is the entry point. Most agents need only this to start working safely on AWS.

2. Skills

Declarative agent capabilities specifically for AWS workflows. Skills are reusable, opinionated patterns that an agent can invoke — think of them as “playbooks” rather than tool calls.

Examples (per AWS launch docs):

“Deploy a serverless API to Lambda + API Gateway.”
“Set up a CloudFront + S3 static site.”
“Configure DynamoDB single-table design.”
“Set up CDK project with proper IAM scoping.”
“Deploy a Bedrock-backed agent.”

Skills are agent-agnostic — they work in Claude Code (via Anthropic’s Skills system), Cursor, Codex, and others. AWS ships them so customers don’t reinvent the same workflows.

3. Plugins

Agent-specific drop-in extensions. Plugins integrate the AWS MCP Server + skills into specific agent UIs:

Claude Code plugin — installs the AWS MCP Server with one command and pre-wires AWS skills.
Cursor plugin — adds AWS MCP to the Cursor MCP marketplace, with native UI for AWS resources.
Codex plugin — integrates with Codex CLI / desktop / VS Code extension.
Kiro / Kiro CLI — pre-installed natively.
Cline / Windsurf plugins — community-maintained but AWS-blessed.

Plugins are how customers actually feel the toolkit — they hide the MCP wiring.

How the toolkit fits in the broader ecosystem

The Agent Toolkit sits between two layers:

Above: AI coding agents (Claude Code, Cursor, Codex, Kiro, Cline, Windsurf, ChatGPT Atlas, OpenClaw, JetBrains Air).
Below: AWS services (EC2, S3, Lambda, Bedrock, CloudFormation, DynamoDB, etc.).

The toolkit is the governed bridge. Without it, agents either hit AWS via raw API keys (bad governance) or the developer manually installs a community MCP server (works but less audit-friendly). The Agent Toolkit makes the safe path the default path.

Why AWS shipped the Agent Toolkit now

Three forces converged in early 2026:

MCP became the standard. The Linux Foundation’s Agentic AI Foundation took governance of MCP, with 146+ member organizations. AWS, Microsoft, and Google could safely ship first-class MCP servers without proprietary lock-in concerns.
Agentic AI hit production. Bedrock Managed Agents (powered by OpenAI), Microsoft Agent 365, Salesforce Agentforce 360, and Anthropic Claude Code all went GA in early 2026 — making “agent governance on AWS” a real customer ask.
Phantom AI Work risk. Enterprises started auditing AI-driven actions; AWS needed first-class IAM + audit for agent-driven AWS calls, not just human-driven ones.

The Agent Toolkit is AWS’s structural answer.

What the AWS MCP Server actually adds vs raw API keys

Concern	Raw API keys	AWS MCP Server
Authentication	Static or rotated keys	IAM identity (per-agent role)
Authorization	IAM policy on key	IAM policy with MCP context keys
Audit	CloudTrail (mixed with human calls)	CloudTrail with `AWS-MCP` separation
Rate limiting / scoping	Manual	Built-in approval gates
Read-only safety	Manual policy	Default-on read-only mode
Metrics	Manual CloudWatch setup	Pre-configured `AWS-MCP` namespace

For any production AWS workload involving an AI agent, the AWS MCP Server is now the default-correct integration. Raw API keys are the Phantom-AI-Work-creating anti-pattern.

How to use the Agent Toolkit

A typical install for a Claude Code user wanting AWS access:

# Install AWS MCP Server in Claude Code
claude mcp add aws-mcp-server

# Authenticate with your AWS account (uses standard AWS SDK credential chain)
aws sso login --profile prod

# Now Claude Code can call AWS services with your IAM identity
# CloudTrail logs every call, CloudWatch publishes metrics under AWS-MCP namespace

For Cursor:

// In .cursor/mcp.json
{
  "mcpServers": {
    "aws": {
      "command": "uvx",
      "args": ["aws-mcp-server"],
      "env": { "AWS_PROFILE": "prod" }
    }
  }
}

The pattern is similar across Codex, Cline, Windsurf, and others — the MCP layer is consistent.

Where the Agent Toolkit fits vs alternatives

Need	Use
Any AWS service access from an agent	AWS MCP Server (Agent Toolkit)
Deep AWS service operations	awslabs/mcp specialized servers
Microsoft 365 + Azure work	Microsoft Agent 365, Power Platform connectors, Microsoft MCP servers
GCP-native work	Google Workspace Studio, Google’s MCP servers (announced at Cloud Next 2026)
Salesforce / SaaS	Vendor-official MCPs (Salesforce, GitHub, Atlassian, Slack)
Self-hosted / VPC	Coder Agents (May 6, 2026) with MCP support

The Agent Toolkit is the AWS slice. Other clouds and vendors are shipping their own equivalents.

What’s still maturing

Honest gaps as of May 7, 2026:

Skills catalog is small. AWS shipped the toolkit with a starter set of skills; community + AWS will grow it over the next 6-12 months.
Bedrock-deep skills are partial. Bedrock Knowledge Bases, AgentCore, and Bedrock Managed Agents have separate setups; Agent Toolkit doesn’t fully unify them yet.
Multi-region setups need configuration. The default works in one region; cross-region agent flows take config.
Per-agent IAM roles are easier in theory than practice. Most teams start with shared IAM and gradually split.

These are normal launch-month gaps. Watch for AWS to publish more skills and plugins through re:Invent 2026.

Bottom line

The AWS Agent Toolkit is the right default for any AI coding agent that needs to safely build on AWS in May 2026. It bundles the new GA AWS MCP Server, declarative skills, and agent-specific plugins into one official AWS-supported package — making the “safe path” the default path. Combined with the existing awslabs/mcp specialized servers and the broader MCP ecosystem (vendor-official MCPs for non-AWS services), the toolkit gives enterprises a governed, audited way to let agents build on AWS without reinventing the integration. Expect every serious AWS-native AI coding workflow to standardize on it within 6 months.

Sources: AWS Blog “The AWS MCP Server is now generally available” (May 6, 2026), aws/agent-toolkit-for-aws GitHub repository (May 2026), AWS Agent Toolkit product page (May 2026), AWS What’s Next with AWS 2026 announcements (April 28, 2026), Toolradar Best MCP Servers 2026 (May 2026).