Best AWS MCP Servers in 2026: 7 Tools That Turn Claude into a Cloud Engineer

Table of Contents

1. What Is an AWS MCP Server?
2. Why AWS + MCP Now
3. How We Ranked These
4. The 7 Best AWS MCP Servers
5. Quick Comparison Table
6. Setup: Wiring AWS Credentials Safely
7. Workflows Worth the Setup
8. Cost Controls You Must Set
9. Frequently Asked Questions

What Is an AWS MCP Server?

An AWS MCP server is a small program that runs locally (or on a server) and exposes one or more AWS services to an AI client over the Model Context Protocol. Some servers focus on a single service — S3, DynamoDB, Lambda. Others are generalists that proxy a wide range of AWS APIs through a unified tool surface. The AI client sees them as a set of tools it can invoke; you see Claude answering questions about your cloud that would otherwise require the AWS console.

For deep configuration walkthroughs read our companion AWS MCP servers setup guide and the focused AWS Docs MCP review.

Why AWS + MCP Now

Three forces have made AWS MCPs the breakout category of 2026:

• AWS Bedrock and Claude. Anthropic's models are first-class on Bedrock, which means a growing number of AWS-native teams want their AI workflow inside AWS. MCP is the wiring that makes that practical.
• Cost pressure. FinOps teams need a way to ask "where did the bill go this month?" without writing custom dashboards. An AWS Cost Explorer MCP plus Claude answers that in 30 seconds.
• Incident response. When CloudWatch fires at 3 AM, on-call engineers want an assistant that can read the alarm, pull logs, check IAM roles, and propose a fix — not a chatbot that just talks about cloud architecture.

How We Ranked These

We scored every AWS-related MCP in the directory on five signals: service coverage (how many AWS APIs are exposed and how cleanly), security posture (does it default to read-only? does it warn on destructive operations?), GitHub stars + commit recency, README quality, and stability under real load. The top 7 below score highest on the composite.

The 7 Best AWS MCP Servers

1. AWS Docs MCP Server

The safest, most useful MCP to start with. AWS Docs MCP gives Claude grounded access to AWS official documentation — service references, API reference, best-practice guides, and code samples. No AWS account required, no IAM credentials to manage, no risk of unexpected charges.

Best for: Anyone learning AWS, architecture sessions where you need cited answers, teams who want documentation lookup without leaving the chat.

Install: npx -y @smithery/aws-docs

2. rafalwilinski AWS MCP

The most-starred general-purpose AWS MCP in the registry. Wraps the AWS SDK and exposes a broad slice of commonly used services — S3, DynamoDB, Lambda, CloudWatch, IAM, EC2. Designed for "ask Claude anything about your AWS account" workflows.

Best for: Day-to-day cloud ops, exploration of unfamiliar accounts, anyone who wants one MCP that does most of AWS.

Install: npx -y rafalwilinski-aws-mcp (consult the README for the latest package name and required env vars).

3. AWS Builder MCP

A builder-focused MCP that combines documentation lookup, code generation, and limited execution. The headline feature: Claude can scaffold CDK or Terraform stacks for common architectures, run a dry-plan against your account, and explain the diff before you apply.

Best for: Infrastructure-as-code authoring, junior-engineer onboarding, prototype-to-production scaffolding.

Install: Check the project README — the install command depends on which AWS SDK version you bind to.

4. AWS S3 MCP

A focused server that does S3 well: bucket lifecycle, object CRUD, signed URLs, replication checks, inventory queries. Faster and safer than running the general-purpose MCP when your workflow is S3-heavy.

Best for: Data engineers, anyone running large object stores, backup and restore automation.

Install: Check the project README — package name may differ between forks.

5. DynamoDB Read-Only MCP

A deliberately read-only DynamoDB MCP. Ideal first install for data exploration: query tables, scan partitions, inspect indexes, and check throughput without any risk of accidental writes. Pairs cleanly with a separate write-capable MCP when you need it.

Best for: Data exploration, support engineers investigating customer data, anyone who wants strict guardrails.

Install: npx -y dynamo-readonly-mcp

6. AWS Advisor MCP

Surfaces AWS Trusted Advisor and Cost Optimizer findings through MCP. Useful for FinOps work: ask "what should I tag for FY26 reviews?" or "where am I overprovisioned?" and get a structured list back.

Best for: FinOps teams, monthly cost reviews, AWS Well-Architected reviews.

Install: Check the project README — requires Trusted Advisor entitlements on your account.

7. AWS Serverless MCP

Covers the serverless slice of AWS — Lambda invocations, API Gateway routes, SQS/SNS queues, EventBridge rules, Step Functions executions. Designed for teams whose entire stack lives in Lambda + DynamoDB + S3.

Best for: Lambda-first teams, event-driven architectures, incident response on serverless workloads.

Install: Check the project README — install command varies by version.

Quick Comparison Table

Setup: Wiring AWS Credentials Safely

The single most important rule when installing any AWS MCP: start with a dedicated IAM user that has read-only policies. Do not point the MCP at your developer credentials, and never at root account credentials. Here is the canonical setup:

1. Create an IAM user named claude-mcp-readonly. Attach the AWS-managed policy ReadOnlyAccess (or a tighter custom policy if you want to be strict).
2. Generate an access key. Store it via aws configure --profile claude-mcp so it lives in ~/.aws/credentials rather than in your MCP config file.
3. Point the MCP at the profile using the AWS_PROFILE=claude-mcp environment variable in your MCP config:

{   "mcpServers": {     "aws": {       "command": "npx",       "args": ["-y", "rafalwilinski-aws-mcp"],       "env": { "AWS_PROFILE": "claude-mcp", "AWS_REGION": "us-east-1" }     }   } }

1. Restart your MCP client. The AWS MCP appears in your tool list. Confirm it works by asking Claude to list your S3 buckets — a safe read operation.
2. Add write access later, scoped narrowly. When you want Claude to be able to create resources, add a second IAM user with policies for only the specific services and actions you trust. Never grant :.

The Skiln Config Generator ships AWS-aware presets that include the IAM-profile-binding pattern by default.

Workflows Worth the Setup

The patterns that justify the install effort:

• Account audit on a fresh engagement. Ask Claude to list every active EC2 instance, every public S3 bucket, every IAM user without MFA, and every region with running spend. You get an audit report in two minutes that used to take a half day.
• Cost spike triage. When the monthly bill jumps, Claude can pull the top 10 cost-by-service deltas from Cost Explorer, correlate with recent CloudFormation deploys, and identify the offending change.
• Production incident response. A Lambda is timing out. Claude reads the recent invocations, pulls the CloudWatch logs, identifies the slow downstream call, and proposes a fix.
• Data exploration before writing a query. Ask "how many items in the orders table have status='pending' older than 7 days?" and Claude runs a Query (not a Scan) and returns the count with the partition key it chose.
• Onboarding new engineers. Give them an MCP-enabled Claude instead of a 90-minute AWS console walkthrough. They learn faster, ask better questions, and ramp in days instead of weeks.

For full setup walkthroughs across multiple AWS services, the desktop MCP installation guide covers the AWS family in detail.

Cost Controls You Must Set

Before you ever ask Claude to touch a paid AWS resource, set three guardrails:

1. CloudWatch billing alarm. Set a hard threshold at 25% above your normal monthly baseline. The alarm gives you a 24-hour head start on any runaway query.
2. Budget rules. Add an AWS Budgets rule that emails on 50%, 80%, and 100% of forecasted spend.
3. Service Control Policies (for orgs). If you use AWS Organizations, attach an SCP that denies high-risk actions (e.g., creating expensive instance types, opening large RDS instances) on the MCP-bound account.

If you hit common errors during setup, the MCP installation errors guide covers the AWS-specific failure modes.

Frequently Asked Questions

What is an AWS MCP server?

An AWS MCP server is a Model Context Protocol server that connects AI clients (Claude Desktop, Claude Code, Cursor) to one or more AWS services — S3, DynamoDB, Lambda, CloudWatch, IAM, EC2, and others. The MCP exposes service-specific tools that Claude can call, so you can ask 'list my S3 buckets' or 'show me the last 100 Lambda invocations' in plain English and get a real answer.

Is there one MCP that covers all of AWS?

Not really. AWS is a sprawling platform — 200+ services, each with its own API surface. In practice, you compose two or three MCPs that cover the services you actually use: AWS Docs MCP for reference lookups, a general AWS MCP like rafalwilinski for high-level operations, plus service-specific MCPs (S3, DynamoDB, Lambda) when you need depth. We list every option in the Skiln directory.

How do AWS MCP servers handle credentials?

Every reputable AWS MCP reads from the standard AWS credentials chain: environment variables (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY), the shared credentials file (~/.aws/credentials), or IAM roles when running on EC2. Best practice is to use a dedicated IAM user with read-only policies for exploration MCPs, and to scope write-capable MCPs to specific services via narrowly defined IAM policies.

Will an AWS MCP rack up my bill?

It can — but the risk is mostly from inefficient queries (e.g., asking Claude to scan a 1 TB DynamoDB table when a query would suffice). Read operations on most services are cheap or free. The danger zones are unbounded list operations, full table scans, and large data egress. Set CloudWatch billing alarms and use read-only IAM credentials for the first 30 days while you learn what Claude calls.

Which AWS MCP should I install first?

Start with the AWS Docs MCP — it has no AWS account requirement, no risk of unexpected charges, and gives Claude grounded answers from Amazon's official documentation. Once you are comfortable with how MCPs work, layer on a read-only general-purpose AWS MCP, then add service-specific MCPs for the services you use most.

Do AWS MCP servers support multiple accounts or profiles?

Yes. Every server in this guide respects AWS profile selection via the AWS_PROFILE environment variable or the --profile flag. For multi-account workflows, point the MCP at a profile that uses sts:AssumeRole to switch into the target account on demand. Avoid hardcoding long-lived root credentials.

Can I run an AWS MCP from Claude Code on a remote EC2 instance?

Yes. Claude Code can SSH into a remote host that runs the MCP server with an attached IAM instance role. The instance role provides credentials automatically and removes the need to copy access keys around. This is the recommended pattern for production-grade ops workflows.

Where can I see every AWS MCP indexed on Skiln?

Visit /browse and search for 'aws' to see every AWS-related MCP, skill, agent, and command in the directory — ranked by stars, source registry, and active maintenance.

Last updated: June 05, 2026 · Skiln tracks new AWS MCP releases daily across PulseMCP, Smithery, Glama, LobeHub, and mcp.directory.