AWS MCP Servers: The Complete Guide to AWS's Official Directory

What Are AWS MCP Servers?

Model Context Protocol (MCP) is the open standard, originally published by Anthropic, that lets AI models like Claude call external tools and read external data through a consistent interface. An AWS MCP server is simply an MCP server built to expose a specific AWS capability — querying DynamoDB, authoring a CloudFormation template, invoking a Lambda function, or searching AWS documentation — as a set of tools and resources an AI client can call safely.

AWS Labs, the open-source engineering group inside AWS, publishes and maintains the reference implementations in a single GitHub repository: awslabs/mcp. It is free, Apache-2.0 licensed, and has accumulated more than 9,500 GitHub stars and 1,700 forks as engineering teams standardize on it for AI-assisted cloud operations.

The Official awslabs/mcp Directory

Unlike community-maintained lists for other clouds, AWS consolidates its MCP servers in one actively maintained monorepo. Each server ships as an independent Python package (installable via uvx or pip), documented on the companion site awslabs.github.io/mcp.

Two servers stand out as the best entry points for most teams:

AWS MCP Servers by Category

The directory spans six practical categories. Search or filter the full list below, and open any server's documentation to see its exact tool set and installation command.

Bedrock AgentCore Gateway: Turning APIs Into MCP Tools

The awslabs/mcp servers cover AWS services themselves, but most real agent workflows also need your own internal APIs, Lambda functions, and SaaS tools exposed as MCP tools. That is the job of Amazon Bedrock AgentCore Gateway, a fully managed AI gateway that has been generally available since October 2025.

Rather than acting only as an MCP passthrough, AgentCore Gateway is a single secure entry point for all agentic traffic. It converts OpenAPI specs, Smithy models, and Lambda functions into MCP-compatible tools, fronts other agents through Agent-to-Agent (A2A) passthrough targets, and routes inference requests across model providers.

Core Gateway Capabilities

• 1-click connectors for Salesforce, Slack, Jira, Asana, and Zendesk without writing custom integration code
• Framework compatibility with CrewAI, LangGraph, LlamaIndex, and Strands Agents
• Serverless operation — no infrastructure to provision, patch, or scale
• MCP 2026-07-28 support — the protocol's largest revision to date, adding a stateless architecture and governed extensions system

AgentCore Gateway Pricing

AgentCore Gateway bills separately from the tools it fronts. Here is the pricing that actually matters when you are estimating cost for an agentic-tool project:

Tool execution itself — a Lambda invocation, a third-party API call, a database query — is billed by whatever service actually runs it, separate from the gateway fee. See the full breakdown on the AWS Bedrock AgentCore Gateway documentation.

AWS MCP Gateway vs Azure and Google Cloud

AWS is not the only hyperscaler shipping managed MCP infrastructure in 2026, but the three clouds have taken noticeably different approaches:

Sources: Microsoft Azure API Management MCP announcement, Apigene MCP gateway cloud comparison.

The practical takeaway: if your stack is already on AWS, AgentCore Gateway is the lowest-friction path to exposing internal tools to agents — you get managed OAuth, semantic tool discovery, and per-call pricing with no infrastructure to run. Teams on Google Cloud today are more likely to be assembling their own gateway or leaning on the well-adopted MCP Toolbox for Databases rather than a single managed product.

How to Connect Claude to AWS MCP Servers

The fastest path is the AWS Knowledge MCP Server, since it needs no local install and no AWS credentials. Add this block to your MCP client's configuration file (for Claude Desktop, that is claude_desktop_config.json):

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

For servers that operate on your own AWS account — DynamoDB, Lambda, EKS, and similar — you will install the Python package locally and supply AWS credentials via your existing IAM setup:

1. 1
   Pick your servers from the directory above based on the AWS services your workflow touches.
2. 2
   Install via uvx or pip, following the exact command listed on each server's documentation page.
3. 3
   Scope an IAM role with least-privilege permissions for exactly what that server needs to read or change.
4. 4
   Add the server block to your client config and restart the client to pick up the new tools.
5. 5
   Test with a read-only prompt first before granting write access to production resources.

Choosing the Right AWS MCP Servers

With 40+ servers available, most teams only need a handful. Work through these questions before installing anything:

Security Best Practices for AWS MCP

An MCP server is only as safe as the credentials behind it. Because these servers can read production data and, in some cases, modify infrastructure, treat every connection with the same rigor as a new IAM principal.

• Least privilege — scope IAM roles to exactly the actions each server needs, nothing broader
• Separate credentials per environment — never point a development agent at production account keys
• Enable AgentCore Gateway's dual OAuth so both agent identity and tool credentials are verified independently
• Log and audit tool calls using CloudTrail so every action an agent takes is traceable after the fact

For a deeper look at securing agent-to-tool connections generally, see our guide to MCP security best practices.

Frequently Asked Questions