|
I’ve been constructing with AI brokers and MCP instruments for some time now, and one query stored developing: how do you give an agent actual, authenticated entry to AWS with out handing it the keys to the dominion? In the present day, there’s a solution.
I’m comfortable to announce the overall availability of the AWS MCP Server, a managed distant Mannequin Context Protocol (MCP) server that offers AI brokers and coding assistants safe, authenticated entry to all AWS providers by way of a small, mounted set of instruments.
The AWS MCP Server is a part of the Agent Toolkit for AWS, a set of tooling that features the MCP Server, abilities, and plugins that assist coding brokers construct extra successfully and effectively on AWS.
AI coding brokers are already helpful for a lot of duties, however they run into actual bother when working with AWS at any significant depth. With out entry to present AWS documentation, brokers depend on coaching information that could be months old-fashioned and will not find out about providers like Amazon S3 Vectors, Amazon Aurora DSQL, or Amazon Bedrock AgentCore. When requested to construct infrastructure, they have an inclination to achieve for the AWS Command Line Interface (AWS CLI) slightly than AWS Cloud Growth Package (AWS CDK) or AWS CloudFormation, and so they produce AWS Identification and Entry Administration (IAM) insurance policies which might be far broader than vital. The result’s infrastructure that works in a demo however will not be production-ready.
The AWS MCP Server addresses this by way of a compact set of instruments that don’t devour your mannequin’s context window. The call_aws device executes any of the 15,000+ AWS API operations utilizing your current IAM credentials. Once we will launch new APIs, they are going to be supported inside days. The search_documentation and read_documentation instruments retrieve present AWS documentation and greatest practices at question time, so the agent all the time works from up-to-date info.
With normal availability, we’re introducing a number of new capabilities. The AWS MCP Server now helps IAM context keys, so that you now not want a separate IAM permission to make use of the server and might categorical fine-grained entry in an ordinary IAM coverage. Documentation retrieval now not requires authentication. We’ve got additionally diminished the variety of tokens required per interplay, which issues for complicated, multi-step workflows.
Additionally new, the run_script device lets the agent write a brief Python script that runs server-side in a sandboxed surroundings. The sandbox inherits your IAM permissions however has no community entry, so that you may give an agent the flexibility to course of information with out giving it entry to your native file system or a shell. When an agent must name a number of APIs and mix the outcomes, making them separately is gradual and burns context. With run_script, the agent chains API calls, filters responses, and computes leads to a single round-trip, which is each quicker and extra context-efficient.
Essentially the most important addition is the transition from Agent SOPs to Abilities. Abilities present curated steering and greatest practices for the duties the place brokers mostly make errors. This helps brokers full work quicker, utilizing validated greatest practices, with fewer errors and fewer tokens — all of which saves you money and time. Abilities are contributed and maintained by AWS service groups. This retains the device listing brief and predictable, which reduces hallucination and retains the agent centered.
For enterprise prospects, the AWS MCP Server offers a transparent separation between human and agent permissions. You should utilize IAM insurance policies or Service Management Insurance policies to specify {that a} given consumer can carry out mutating operations whereas the MCP server is restricted to read-only actions. Amazon CloudWatch metrics printed beneath the AWS-MCP namespace allow you to observe MCP server calls individually from direct human calls, supplying you with the audit path that compliance groups require. Amazon CloudTrail captures all API calls for an entire document.
Let’s see it in motion
For this demo, I selected to make use of Claude Code, however I can use the AWS MCP Server with any AI agent that helps MCP, which is mainly all instruments accessible at the moment: Kiro CLI, Kiro, Cursor, Codex, and extra. I configure Claude Code to make use of the Anthropic Opus 4.6 mannequin.
Opus 4.6 has a data cutoff date in Could 2025. It means it doesn’t know something that occurred after Could final 12 months. I ask a query about an AWS service that was launched not too long ago: Amazon S3 Vectors, launched in preview in July 2025 and that went GA in December 2025.
The query is “ retailer embedding on S3″. (embedding is a sort of vector)
It offers me 5 options, all right, however none utilizing S3 Vectors as I requested. Notice that this reply comes from the Opus 4.6 mannequin, not from Claude Code. Any AI device utilizing the identical mannequin will return related solutions as a result of S3 Vectors wasn’t introduced on the time the mannequin was educated.
Let’s now strive with the AWS MCP Server.
The AWS MCP Server makes use of AWS Identification and Entry Administration (IAM) and IAM SigV4 authentication. To make use of my native AWS credentials configuration over MCP, which solely helps OAuth 2.1, I configure my AI coding agent to name the AWS MCP Server by way of a proxy. The MCP Proxy for AWS is an open supply proxy that runs on my machine and bridges the world of IAM authentication to OAuth.
I add the MCP configuration with this command:
claude mcp add-json aws-mcp --scope consumer
'{"command":"uvx","args":["mcp-proxy-for-aws@latest","https://aws-mcp.us-east-1.api.aws/mcp","--metadata","AWS_REGION=us-west-2"]}'
Let’s analyze the JSON configuration:
- I exploit the consumer scope to make the server accessible to all my initiatives on my laptop computer.
uvx mcp-proxy-for-awsis the command to launch the proxy; the remainder of the arguments are parameters handed to the proxy.https://aws-mcp.us-east-1.api.aws/mcpis likely one of the two regional endpoints for the AWS MCP Server. The proxy will ahead Claude Code’s requests to that endpoint.--metadataare handed to the proxy goal. Right here, it tells the AWS MCP Server to make use of the US West (Oregon) Area.
I begin Claude Code and I kind /mcp to confirm the AWS MCP Server is appropriately put in and might use my credentials.
I ask the identical query: “how can I retailer embedding on S3”.
This time, Claude Code is aware of it has a device it will possibly use to reply the query. It asks me permission to invoke the aws___search_documentation device. After a number of seconds, I obtain an accurate reply: “AWS now has a devoted service for this: Amazon S3 Vectors …”
Pricing and availability
The AWS MCP Server is offered at the moment within the US East (N. Virginia) and Europe (Frankfurt) AWS Areas and might make API calls to any Area. There isn’t a extra cost for the AWS MCP server itself. You pay just for the AWS assets you create and any relevant information switch prices.
The AWS MCP Server works with Claude Code, Kiro, Cursor, and any MCP-compatible shopper. To get began, see the AWS MCP Server Person Information.
I’ve been ready for one thing like this since I began utilizing MCP instruments in my AI brokers early final 12 months. The mix of present documentation, authenticated API entry, and sandboxed script execution in a single server adjustments what an agent can truly do on AWS. I’m curious what you construct with it. Let me know within the feedback.