What is the AWS MCP Server and how do you set it up safely?

Short answer

The AWS MCP Server is a managed, remote MCP server run by AWS, generally available since May 6, 2026. It gives AI agents such as Claude Code, Cursor, Codex and Kiro a small set of tools: call any of 15,000+ AWS API operations, search and read AWS documentation, and run short Python scripts in a sandbox. It acts with your IAM permissions and costs nothing extra.

What the server actually exposes

The AWS MCP Server does not wrap each AWS service in its own tool. AWS chose a small, fixed set of tools that cover everything, so the AI’s context does not fill up with hundreds of tool descriptions. According to the GA announcement of May 6, 2026:

Tool What it does
call_aws Runs any of the 15,000+ AWS API operations with your existing IAM credentials. AWS says new APIs are supported within days of launch.
search_documentation and read_documentation Fetch current AWS documentation and best practices at the moment the agent needs them. No authentication needed since GA.
run_script Runs a short Python script server-side in a sandbox. It inherits your IAM permissions but has no network access, no shell and no access to your local files.

Next to the tools, the server offers Skills: curated guidance written by AWS service teams for mistakes agents commonly make. These replaced the earlier “Agent SOPs”.

The documentation tools matter more than they look. A model has a knowledge cutoff, and AWS ships new services and API changes every week. In AWS’s own demo, Claude Code without the server did not know about a service that had launched after its cutoff; with the server it searched the docs and answered correctly.

Setting it up, step by step

  1. Pick the region endpoint. The server runs in US East (N. Virginia), for example https://aws-mcp.us-east-1.api.aws/mcp, and in Europe (Frankfurt). Teams in Europe usually pick Frankfurt. The server can still call APIs in any region.
  2. Install uv on each machine, because the proxy is started with uvx.
  3. Add the server through the MCP Proxy for AWS (mcp-proxy-for-aws). Your MCP client talks OAuth; AWS wants IAM with SigV4 signing. The proxy runs locally and translates. AWS names Claude Code, Kiro, Kiro CLI, Cursor and Codex as tested clients, and says any MCP-compatible client works.
  4. Give the agent its own permissions. Write an IAM policy, or a Service Control Policy at organization level, that limits what calls through the MCP server may do. Start read-only.
  5. Watch it. Calls through the server show up in AWS CloudTrail, and CloudWatch has an AWS-MCP namespace that separates agent calls from calls people make directly.

The common misunderstanding: “the agent only does what I ask”

The agent does what your IAM identity allows. If a developer’s profile can delete an S3 bucket, an agent working through that profile can too, after one misread instruction. The server does not add a safety layer on top of IAM; it makes IAM the safety layer. That is good design, but only if your IAM policies were already tight.

So the real setup work is not the MCP config, it is deciding which role the agent runs under. The GA release made that easier: IAM context keys let you write conditions that apply only when a request comes through the MCP server, so the same person can have broader rights in the console than their agent has. Read is MCP safe to use in a company for the general risks, including prompt injection through data the agent reads.

Managed server or the awslabs servers?

AWS also maintains a large collection of open source MCP servers under awslabs, one per service or workflow, which you run yourself. The managed server is one remote endpoint that covers all APIs through call_aws. For most teams the managed server is the simpler default: nothing to update, one place to set permissions, one audit trail. A dedicated awslabs server can still make sense when you want a narrow, purpose-built tool for one service. The difference between running a server yourself and connecting to a hosted one is explained in local vs remote MCP server.

What this means for a team

Every developer who connects gets the same AWS reach and the same current documentation, whichever AI tool they use. What does not travel is the knowledge about your own account: which stacks are legacy, which bucket holds production data, why a security group is open, what went wrong in last month’s incident. The AWS docs cannot know that, and each person’s agent learns it again in its own session, or not at all. For the general pattern, see how to give AI agents your company context.

Where Bombos fits

The AWS MCP Server gives every agent the same access to AWS; Bombos gives those agents the same memory of how your team runs AWS. It is a plugin underneath Claude Code, Codex, Cursor, Copilot and the other AI tools your team uses that learns how your team works, such as “never touch the prod account from a laptop”, and shows with every answer what it rests on. Your agents act in AWS; Bombos remembers. It runs in the cloud in Europe or standalone on your own server, and one engineer can try it first with their own agents.

Related questions

Does the AWS MCP Server cost money?

AWS says there is no additional charge for the server itself. You pay for the AWS resources your agent creates and for applicable data transfer.

Which regions is it available in?

At GA the server runs in US East (N. Virginia) and Europe (Frankfurt). From there it can make API calls to any AWS Region.

Why do I need mcp-proxy-for-aws?

The server authenticates with IAM and SigV4, while MCP clients speak OAuth. The open source MCP Proxy for AWS runs locally, signs requests with your AWS credentials and bridges the two. It is started with uvx, so uv must be installed.

Can I make the agent read-only?

Yes. AWS recommends separating human and agent permissions with IAM policies or Service Control Policies, for example by making the MCP server read-only. With GA, IAM context keys let you write those rules in a standard IAM policy.

Is this the same as the awslabs MCP servers on GitHub?

No. The awslabs repository holds many open source servers you run yourself, one per service or workflow. The AWS MCP Server is one remote server operated by AWS, part of the Agent Toolkit for AWS.

Sources

  1. AWS News Blog: The AWS MCP Server is now generally available (May 6, 2026)
  2. AWS What's New: The AWS MCP Server is now generally available
  3. AWS Labs MCP servers (open source)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.