4 MCP servers that stop my AI agent guessing about AWS

How many times this week did your AI assistant answer from memory instead of checking the docs? The model knows AWS as it was a few months ago. Since then AWS has shipped dozens of new features, raised quotas and changed prices.

Day to day I work with my AI agent, Johnny Cloud. He’s my architect sidekick: he checks the docs, reviews accounts, works out costs and prepares estimates. He has one hard rule: “I don’t know, I don’t guess”. To stick to it, he needs somewhere to get current data from. That’s where MCP servers come in.

In this post I’ll show you the four servers Johnny has permanently plugged in: for documentation, for working in AWS account, for pricing and for client estimates.

What is MCP?

MCP (Model Context Protocol) is an open protocol for connecting external tools and data sources to an AI assistant. Claude, Kiro, Cursor and VS Code all talk to an MCP server in the same way, and the server exposes specific tools such as “search the documentation” or “call an API”.

For us it means one simple thing. The model doesn’t have to guess, because it can check the most recent information. On AWS, where prices, features and regional availability change almost every week, that makes a big difference.

MCP servers are either local (you run them yourself, usually with uvx) or remote (you connect to an HTTPS endpoint).

1. AWS MCP Server: one server for everything

This is my main server. AWS made it generally available in May 2026. It’s managed by AWS, so there’s nothing to install apart from a small proxy for authentication.

What it can do:

• call_aws runs any of the 15,000+ AWS API operations with your IAM permissions. New APIs are supported within days of launch.
• run_script lets the agent write a short Python script that runs server-side in a sandbox. The sandbox inherits your IAM permissions but has no network or local file access, so several API calls can be chained in one step.
• search_documentation and read_documentation pull current AWS documentation at query time.
• Skills: curated guidance from AWS service teams for the tasks where agents most often get things wrong.

Access goes through IAM. Every call shows up in CloudTrail, and CloudWatch has a separate AWS-MCP metrics namespace, so you can check what a human did and what the agent did. You can restrict the agent with a standard IAM policy or an SCP.

The server itself is free; you pay for whatever resources the agent creates.

What Johnny and I use it for: reviewing a client’s account (“Johnny, show me security groups open to 0.0.0.0/0”), quick Lambda or CloudWatch troubleshooting, checking configuration before an audit. Johnny picks the API calls himself, and I get the answer plus a list of what he checked.

A practical tip: start with a read-only role before you give the agent write access. Really. Johnny started on read-only too.

{
  "mcpServers": {
    "aws-mcp": {
      "command": "uvx",
      "args": [
        "mcp-proxy-for-aws@latest",
        "https://aws-mcp.eu-central-1.api.aws/mcp",
        "--metadata", "AWS_REGION=eu-central-1"
      ]
    }
  }
}

2. AWS Knowledge MCP Server: when knowledge is all you need

Knowledge MCP is a remote server with AWS knowledge: documentation, blog posts, What’s New announcements and Well-Architected guidance. It needs no AWS account and no sign-in, just the endpoint https://knowledge-mcp.global.api.aws.

The tools I use most:

• search_documentation and read_documentation to find and read a documentation page,
• get_regional_availability to check whether a service, API or CloudFormation resource is available in a given Region,
• list_regions for the current list of Regions,
• retrieve_skill for ready-made guidance on common tasks.

I treat it as a lighter alternative to the AWS MCP Server. The AWS MCP Server includes the same documentation tools, but when I don’t need account access, Knowledge server is enough.

Johnny goes to it for every question about a quota, a price or regional availability before he answers me. It’s ideal for preparing training, for a quick check such as “is this in Frankfurt yet?”, or for designing an architecture.

{
  "mcpServers": {
    "aws-knowledge": {
      "type": "http",
      "url": "https://knowledge-mcp.global.api.aws"
    }
  }
}

3. AWS Pricing MCP Server: real prices, no guessing

Pricing is the first place where an AI model gets things wrong. It remembers last year’s rate, or the rate from another Region. The AWS Pricing MCP server (open source, from the awslabs repository) queries the AWS Price List API directly.

What it can do:

• get_pricing returns current prices for a service with filters (Region, instance type, OS and so on), while get_pricing_service_codes and get_pricing_service_attributes help you work out what to ask for,
• analyze_cdk_project and analyze_terraform_project scan an IaC project and list the AWS services to price,
• generate_cost_report builds a cost report with unit prices and usage scenarios,
• get_bedrock_patterns suggests architecture and cost patterns for Amazon Bedrock,
• get_price_list_urls gives links to the full price lists in CSV/JSON.

This server needs an AWS account connection: a profile with the pricing:* permission. It doesn’t read your bill or any account data, only public price lists, and Price List API calls are free.

In practice Johnny uses it to check and compare prices (“how much cheaper is this instance in Ireland than in Frankfurt?”) and for a first rough estimate of a project.

I give him the basic inputs and he comes back with a list of services and their costs.

{
  "mcpServers": {
    "aws-pricing": {
      "command": "uvx",
      "args": ["awslabs.aws-pricing-mcp-server@latest"],
      "env": {
        "AWS_PROFILE": "your-profile",
        "AWS_REGION": "eu-central-1"
      }
    }
  }
}

4. AWS Pricing Calculator MCP: an estimate with a link for the client

Pricing MCP gives you prices, but the client wants to see an estimate in the AWS Pricing Calculator. That’s where sample-aws-pricing-calculator-mcp from aws-samples comes in.

It builds a calculator estimate from a plain-language description and returns a calculator.aws link you can send straight on. It doesn’t need any AWS credentials.

The main tools:

• search_services and get_service_fields find a service and the fields it needs,
• create_estimate, add_service and validate_estimate build the estimate step by step and check the configuration,
• build_estimate does it all in one call,
• export_estimate saves the estimate to calculator.aws and returns the link,
• import_estimate loads an existing estimate from a link, for example to fix it.

A quick note. It’s a sample project from aws-samples, not an official AWS service. It uses an undocumented calculator API that may change. So far 18 of roughly 436 services are verified.

The server doesn’t calculate costs itself; the calculator does that when you open the link. So I open and check every estimate before it goes to a client.

The time saved is real, though. Instead of clicking through a dozen services in the calculator, I describe the architecture to Johnny and get a link to an estimate that just needs a review.

{
  "mcpServers": {
    "aws-pricing-calculator-mcp-server": {
      "command": "npx",
      "args": ["-y", "sample-aws-pricing-calculator-mcp@latest"]
    }
  }
}

How do we use them together?

A typical flow when Johnny and I start a new project:

1. Knowledge MCP to check that services are available in the client’s Region and what the current quotas are.
2. Pricing MCP to compare options and Regions.
3. Calculator MCP for the estimate I send to the client.
4. AWS MCP Server after deployment, for account reviews and troubleshooting.

Neither Johnny nor any of these servers replaces an architect’s judgement. The decisions are still mine. But together they save me a huge amount of time.

If you want to build an agent like this yourself, we cover MCP servers and Strands Agents on the Generative AI on AWS training. And if you need an AI agent in your product, see how I help with agentic system architecture.

Which MCP servers do you use with AWS? I’m curious whether I’ve missed one.

Sources: - The AWS MCP Server is now generally available (AWS News Blog) - AWS MCP Server endpoints and quotas - AWS Knowledge MCP Server now generally available (What’s New) - AWS Pricing MCP Server (awslabs) - sample-aws-pricing-calculator-mcp (aws-samples)