AWS MCP Server Officially Launches
Key point
AWS has officially launched an MCP Server for AI agents.
Details
AWS has officially released the AWS MCP Server. It is a managed remote MCP (Model Context Protocol) server that lets AI agents and coding assistant tools access AWS services broadly, in an authenticated state, using only a small number of fixed tools. It also consumes less context window. It is part of the Agent Toolkit for AWS.
Previously, agents relied on outdated training data without up-to-date documentation, making them prone to falling back on the AWS CLI or creating IAM policies broader than CDK/CloudFormation. To reduce these issues, AWS MCP Server provides the following tools.
call_aws: Executes 15,000+ AWS API operations using existing IAM credentials, with support for new APIs added within days.search_documentation/read_documentation: Looks up the latest AWS documentation and best practices.run_script: Inherits IAM permissions but runs Python scripts in a sandbox without network access or local filesystem/shell access, allowing multiple APIs to be chained together at once along with filtering and computation.
Along with GA, IAM context keys have been added, making it possible to define fine-grained access using standard IAM policies without needing separate server-specific permissions. Documentation lookups can now be done without authentication, and the number of tokens needed per interaction has also decreased. Another change is the switch from the existing Agent SOPs to Skills, allowing use of verified guides maintained by AWS service teams. The goal is to keep the tool list short and predictable to reduce hallucinations and improve task speed and accuracy.
In enterprise settings, IAM policies or SCPs can be used to let humans perform change operations while restricting the MCP server to read-only access. Agent calls can be observed separately from direct human calls via CloudWatch's AWS-MCP namespace, and CloudTrail logs a complete record of all API calls. In a demo, Claude Code and Opus 4.6 missed Amazon S3 Vectors—which went through a July 2025 preview and December GA—due to a May 2025 knowledge cutoff, but once a local MCP Proxy for AWS bridged IAM SigV4 and OAuth, they answered the same question correctly.
The service is currently available in US East (N. Virginia) and Europe (Frankfurt), and API calls can be sent to any region. The server itself incurs no additional charge; only AWS resource usage and data transfer costs are billed separately. It can be used with MCP-compatible clients such as Claude Code, Kiro, Cursor, and Codex.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.