Building Secure AI Agents with the AWS MCP Server

Overview

In the rapidly evolving landscape of generative AI, AI coding agents have transformed how developers write and deploy software. However, when these agents interact with complex cloud environments, they often encounter limitations, such as relying on outdated training data, hallucinating API parameters, or failing to grasp intricate multi-step workflows. To address these critical bottlenecks, Amazon Web Services recently launched the Agent Toolkit for AWS, featuring the fully managed AWS MCP (Model Context Protocol) Server. This production-ready suite equips popular AI coding assistants, including Kiro, Claude Code, and Cursor, with secure, auditable, and real-time access to AWS services, official documentation, and curated best practices, ensuring that your AI agents build on AWS efficiently and securely.

Introduction

The integration of AI into the software development lifecycle has brought undeniable speed, but it has also introduced new challenges. Developers using AI coding agents to build cloud-native applications frequently find that these tools struggle with complex, multi-service architectures. Because large language models are trained on historical data, they often lack awareness of newly released AWS services or updated SDK paradigms. Consequently, agents might choose the wrong service, misconfigure Identity and Access Management (IAM) permissions, or write code that looks plausible but ultimately fails to compile.

To solve this, AWS introduced the Agent Toolkit for AWS, effectively superseding the scattered community tools previously available via AWS Labs. At the heart of this toolkit is the AWS MCP Server, which reached General Availability in May 2026 and received robust OAuth support in July 2026. The toolkit provides a standardized framework using the open-source Model Context Protocol to give AI agents dynamic access to over 15,000 AWS APIs. By providing real-time documentation retrieval and tested procedural workflows, the Agent Toolkit empowers AI assistants to move from simple code generation to autonomous, reliable cloud infrastructure management.

Core Components of the Agent Toolkit for AWS

The Agent Toolkit for AWS is not a single tool but a cohesive suite designed to bridge the gap between AI reasoning and deterministic cloud operations. It consists of four primary components:

1. The AWS MCP Server

The AWS MCP Server is a fully managed, remote server that facilitates secure interactions between AI coding agents and your AWS environment. Instead of consuming the full context window of an LLM with API definitions, the MCP server exposes a compact set of highly capable tools.

• call_aws: Allows agents to execute authenticated API calls across the vast AWS ecosystem, complete with syntax validation and error handling.
• run_script: A breakthrough feature that enables agents to write and execute Python code in a sandboxed, server-side environment. This is ideal for chaining multiple API calls, filtering responses, and computing results in a single round-trip, significantly reducing token consumption and latency.
• search_documentation and read_documentation: Grant agents unauthenticated access to the latest AWS documentation, ensuring they work with up-to-date best practices and avoid hallucinated logic.

2. Agent Skills

Previously known as Agent SOPs (Standard Operating Procedures), Agent Skills are curated, on-demand packages of instructions, code scripts, and reference materials. Authored and maintained by AWS service teams, skills address the specific tasks where AI agents most commonly make mistakes.

For example, in June 2026, AWS released open-source skills for AWS SDK best practices. Agents often write SDK code that fails to compile, such as missing the async-throwing nature of the AWS SDK for Swift. By dynamically loading the aws-sdk-swift-usage skill, the agent gains immediate context on modern Swift concurrency patterns, paginators, and waiters, allowing it to generate flawless, compilable code. Agents load only the skills relevant to their current task, keeping the context window lean and focused.

3. Plugins

To simplify installation, the Agent Toolkit offers Plugins, single-install packages that bundle the AWS MCP Server configuration and a curated set of Agent Skills for specific IDEs like Claude Code and Codex. For instance, installing the aws-core plugin instantly equips your AI assistant with the knowledge to handle service selection, AWS CloudFormation deployments, serverless architectures, and observability configurations. Other specialized plugins include aws-data-analytics for Amazon OpenSearch and AWS Glue workflows, and aws-agents-for-devsecops for vulnerability scanning.
4. Rules Files

Rules files are project-level configurations that establish boundaries and operational preferences for your AI agents. These files instruct the agent on how to engage with the AWS MCP Server, prompting it to discover available skills or consult documentation before executing any mutating infrastructure changes.

Security, Governance, and OAuth Integration

When delegating cloud operations to autonomous agents, security is the paramount concern. Traditional applications follow deterministic code paths, whereas AI agents reason dynamically and are susceptible to prompt injection or hallucinations. The Agent Toolkit for AWS introduces enterprise-grade governance to mitigate these risks.

IAM Context Keys and Differentiation

The AWS MCP Server automatically injects two global condition context keys into every downstream AWS service call: aws:ViaAWSMCPService and aws:CalledViaAWSMCP. This allows cloud administrators to write granular IAM policies that differentiate between human-initiated actions and AI-driven actions. For example, an administrator can enforce a policy that restricts AI agents to read-only operations or limits their deployment capabilities to specific development environments, even if the underlying IAM role possesses broader administrative rights. Furthermore, every API call made through the MCP server is logged in AWS CloudTrail, providing a transparent audit trail for compliance teams.

OAuth Support

In July 2026, AWS significantly enhanced the authentication process by introducing OAuth support for the AWS MCP Server. AI agents can now connect directly using AWS Sign-In with industry-standard OAuth 2.0. This eliminates the need for complex, local credential proxies and allows developers to authorize agents interactively via a browser or programmatically through headless authorization. Administrators maintain complete control using dynamic client registration, token introspection, and revocation APIs, seamlessly bridging AI agent access with existing AWS IAM Identity Center configurations.

Monitoring and Observability

Operating AI agents in production requires deep visibility into their behavior. The AWS MCP Server publishes operational metrics directly to Amazon CloudWatch under the AWS-MCP namespace at no additional cost. Engineering teams can monitor invocation counts, track success and error rates, and set up alarms for API throttling. By analyzing usage patterns for specific tools such as call_aws or search_documentation, organizations can identify permission bottlenecks, optimize prompt engineering, and ensure their AI agents operate within expected operational thresholds.

Conclusion

The Agent Toolkit for AWS marks a definitive shift in how developers leverage AI for cloud architecture. By transitioning from generic code generation to context-aware, skill-driven cloud operations, AWS has solved the persistent issues of LLM hallucinations and outdated knowledge. With the fully managed AWS MCP Server, curated Agent Skills, and rigorous IAM-based security controls such as OAuth and context keys, developers can finally trust AI agents to build, deploy, and troubleshoot complex AWS workloads safely. As organizations accelerate their modernization efforts, the Agent Toolkit for AWS stands out as the essential bridge between the reasoning power of generative AI and the deterministic reality of enterprise cloud infrastructure.

Drop a query if you have any questions regarding Agent Toolkit, and we will get back to you quickly.