aws-mcp-server

MCP.Pizza Chef: alexei-led

Rather than wrapping each Amazon service separately, this offers just two abilities: look up the documentation for any AWS command, and run one, optionally filtered through common text tools. That means it can do anything your own permissions allow, including creating, changing, and deleting real resources that cost real money. A local sandbox limits what a command can touch on your own disk, on Linux and macOS only, but places no limit at all on the AWS side. Your permissions are the real boundary.

Deprecated · project moved to the cloud-mcp-server repository
Coding
Data

Use This MCP server To

Ask which servers are running and roughly what they cost Find out why a storage bucket is not reachable Look up the right command for something I do rarely List everything tagged for one project across regions Check who has access to a particular resource Summarise last month's spending by service

README

AWS MCP Server

CI PyPI Code Coverage Linter: Ruff Docker Image

Give Claude access to all 200+ AWS services through the AWS CLI.

Demo

demo.mp4

What It Does

This MCP server lets Claude run AWS CLI commands on your behalf. Instead of wrapping each AWS API individually, it wraps the CLI itself—giving Claude complete AWS access through just two tools:

Tool Purpose
aws_cli_help Get documentation for any AWS command
aws_cli_pipeline Execute AWS CLI commands with optional pipes (jq, grep, etc.)

Claude learns commands on-demand using --help, then executes them. Your IAM policy controls what it can actually do.

flowchart LR
    Claude[Claude] -->|MCP| Server[AWS MCP Server]
    Server --> CLI[AWS CLI]
    CLI --> AWS[AWS Cloud]
    IAM[Your IAM Policy] -.->|controls| AWS
Loading

What's New

  • Streamable HTTP transport — New streamable-http transport for web-based MCP clients, replacing the deprecated sse transport (#33)
  • Input validation error handling — Validation errors now return proper MCP tool errors (isError: true) instead of regular results (#34)
  • Server description — Server advertises its purpose to MCP clients via the instructions field (#35)
  • Server icons — Server provides icon metadata for MCP client display (#36)
  • Graceful shutdown — Server disconnects cleanly when the MCP client disconnects (#16)

Quick Start

Prerequisites

Claude Code

Add to your MCP settings (Cmd+Shift+P → "Claude: Open MCP Config"):

{
  "mcpServers": {
    "aws": {
      "command": "uvx",
      "args": ["aws-mcp"]
    }
  }
}

Claude Desktop

Add to your Claude Desktop config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "aws": {
      "command": "uvx",
      "args": ["aws-mcp"]
    }
  }
}

Docker (More Secure)

Docker provides stronger isolation by running commands in a container:

{
  "mcpServers": {
    "aws": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "-v",
        "~/.aws:/home/appuser/.aws:ro",
        "ghcr.io/alexei-led/aws-mcp-server:latest"
      ]
    }
  }
}

Note: Replace ~/.aws with the full path on Windows (e.g., C:\Users\YOU\.aws).

Docker with Streamable HTTP Transport

For web-based MCP clients, use the streamable-http transport:

docker run --rm -p 8000:8000 \
  -e AWS_MCP_TRANSPORT=streamable-http \
  -v ~/.aws:/home/appuser/.aws:ro \
  ghcr.io/alexei-led/aws-mcp-server:latest

The server will be available at http://localhost:8000/mcp.

Note: The sse transport is deprecated. Use streamable-http instead.

AWS Credentials

The server uses the standard AWS credential chain. Your credentials are discovered automatically from:

  1. Environment variables: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY
  2. Credentials file: ~/.aws/credentials
  3. Config file: ~/.aws/config (for profiles and region)
  4. IAM role: When running on EC2, ECS, or Lambda

To use a specific profile:

{
  "mcpServers": {
    "aws": {
      "command": "uvx",
      "args": ["aws-mcp"],
      "env": {
        "AWS_PROFILE": "my-profile"
      }
    }
  }
}

Configuration

AWS Settings

Environment Variable Description Default
AWS_PROFILE AWS profile to use default
AWS_REGION AWS region (also accepts AWS_DEFAULT_REGION) us-east-1
AWS_CONFIG_FILE Custom path to AWS config file ~/.aws/config
AWS_SHARED_CREDENTIALS_FILE Custom path to credentials file ~/.aws/credentials

Server Settings

Environment Variable Description Default
AWS_MCP_TIMEOUT Command execution timeout in seconds 300
AWS_MCP_MAX_OUTPUT Maximum output size in characters 100000
AWS_MCP_TRANSPORT Transport protocol (stdio, sse, or streamable-http) stdio
AWS_MCP_SANDBOX Sandbox mode (auto, disabled, required) auto
AWS_MCP_SANDBOX_CREDENTIALS Credential passing (env, aws_config, both) both

Security

Your IAM policy is your security boundary. This server executes whatever AWS commands Claude requests—IAM controls what actually succeeds.

Best practices:

  • Use a least-privilege IAM role (only permissions Claude needs)
  • Never use root credentials
  • Consider Docker for additional host isolation

For detailed security architecture, see Security Documentation.

Documentation

  • Usage Guide — Tools, resources, and prompt templates
  • Security Architecture — IAM + Sandbox + Docker model
  • Development Guide — Contributing and testing

License

MIT License — see LICENSE for details.

aws-mcp-server FAQ

Do I need a key?
Yes, your existing AWS credentials. It reads them from the same place your own AWS command line already looks, so there is nothing new to paste.
Can it cost me money?
Yes. It runs real commands, so it can start servers and create resources that land on your bill. There is no spending guard.
Can it delete things?
Yes, if your own permissions allow it. There is no list of forbidden commands, so whatever you could do, it can do.
The summary mentions a safe container, does that protect my account?
No. The sandbox limits what a command reads and writes on your own computer, and only on Linux and macOS. It restricts nothing on the Amazon side.
Can I use this to review my setup without risk?
Yes, if you point it at read-only credentials. That is the sensible way to start.
Which apps does it work in?
Claude Desktop, Claude Code, Cursor, and others, either as a local program or as a small web service.
How hard is setup?
You need the AWS command line already installed and signed in, then a short settings block. A ready-made container image is also published.
Is it maintained, and is the address still right?
The last code change was February 2026, and the project has since moved to a repository named cloud-mcp-server.