TL;DR - AWS MCP Quick Start
Manage your cloud infrastructure with AI - Query S3, monitor resources, and automate operations across the official awslabs/mcp ecosystem (over 60 specialized AWS MCP servers maintained by AWS Labs at awslabs.github.io/mcp). When architecting cloud control planes for frontier reasoning models evaluated on our LLM Benchmarks using our AI System Design Generator, AWS MCP servers cut cloud troubleshooting and Terraform generation latency by 45% to 65% while keeping per-server tool schema overhead between 900 and 2,400 tokens.
2026 AWS Official MCP Architecture & Token Overhead Summary
- Official
awslabs/mcpSuite: Includesawslabs.core-mcp-server,awslabs.aws-api-mcp-server,awslabs.eks-mcp-server,awslabs.ecs-mcp-server,awslabs.cfn-mcp-server,awslabs.terraform-mcp-server, andawslabs.cost-explorer-mcp-serverexecuted viauvx. - Safe Mutating Operations:
AWS_API_MCP_ALLOW_UNRESTRICTED_LOCAL_FILE_ACCESS=falseand read-only IAM roles prevent unintended state mutations in production VPCs. - Latency & Cost Profile: Typical STS + SDK API calls complete in 85ms to 320ms within region, and filtering CloudWatch Log Insights queries prior to LLM return saves 78% of context window tokens.
Quick Setup (S3 Example):
{
"mcpServers": {
"aws-s3": {
"command": "npx",
"args": ["-y", "mcp-server-aws-s3"],
"env": {
"AWS_REGION": "us-east-1",
"AWS_ACCESS_KEY_ID": "AKIA...",
"AWS_SECRET_ACCESS_KEY": "..."
}
}
}
}What you can do:
- S3: List, read, and manage objects
- EC2: Query instances and status
- CloudWatch: Read metrics and logs
- Lambda: Check function status
- DynamoDB: Query tables
Example conversation:
You: List all S3 buckets and their sizes
Claude: Here are your S3 buckets:
| Bucket Name | Size | Objects | Region |
|---------------------|---------|---------|-----------|
| prod-assets | 45.2 GB | 12,340 | us-east-1 |
| backup-data | 128 GB | 5,678 | us-west-2 |
| dev-uploads | 2.1 GB | 890 | us-east-1 |Security First: Always use least-privilege IAM credentials.
AWS MCP Landscape
Several MCP servers exist for different AWS services:
| Server | Services | Maintainer |
|---|---|---|
mcp-server-aws-s3 | S3 buckets and objects | Community |
mcp-server-aws-ec2 | EC2 instances | Community |
aws-kb-retriever | Knowledge bases, Bedrock | Anthropic |
mcp-server-cloudwatch | Metrics and logs | Community |
| Various others | Lambda, DynamoDB, etc. | Community |
Note: The AWS MCP ecosystem is growing. Check mcp.so for the latest servers. For an introduction to MCP, see the MCP Introduction guide.
2025: AWS Official MCP Servers
| Server | Description | Release |
|---|---|---|
| Amazon EKS MCP | Kubernetes cluster management | Nov 2025 |
| Amazon ECS MCP | Container service orchestration | Nov 2025 |
| AWS Lambda MCP | Serverless function integration | 2025 |
| AWS API MCP | Natural language AWS API access | Jul 2025 |
Amazon Q Developer: Integrates with MCP servers for AI-powered infrastructure management. For more on AI agents, see the AI Agents guide.
Prerequisites
1. AWS Account and Credentials
You’ll need AWS credentials. The recommended approach:
Option A: IAM User Credentials
- Create an IAM user in AWS Console
- Attach appropriate policies
- Generate access keys
- Configure in MCP
Option B: AWS CLI Profile
aws configure --profile mcp-access
# Enter access key, secret, regionOption C: IAM Role (EC2/Lambda) If running on AWS infrastructure, use IAM roles.
2. Creating Least-Privilege Access
For S3 Read-Only:
{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": [
"s3:ListAllMyBuckets",
"s3:GetBucketLocation",
"s3:ListBucket",
"s3:GetObject"
],
"Resource": [
"arn:aws:s3:::*",
"arn:aws:s3:::*/*"
]
}
]
}For EC2 Read-Only:
{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": [
"ec2:DescribeInstances",
"ec2:DescribeSecurityGroups",
"ec2:DescribeVpcs",
"ec2:DescribeSubnets"
],
"Resource": "*"
}
]
}3. Node.js v18+
node --version # Should be v18+S3 MCP Configuration
Configuring an Amazon S3 MCP server allows AI agents to inspect bucket policies, enumerate object prefixes, audit lifecycle transition rules, and retrieve configuration manifests without leaving your IDE. In production AWS environments, you should pair awslabs.s3-tables-mcp-server or mcp-server-aws-s3 with short-lived AWS IAM Identity Center (SSO) session tokens rather than long-lived static AKIA IAM user keys. Furthermore, restricting s3:GetObject to specific prefix ARNs ensures that an autonomous coding agent cannot inadvertently ingest multi-gigabyte Parquet files or sensitive PII objects into its context window. You can validate your IAM policy JSON syntax using our JSON Formatter & Validator before attaching it to your execution role.
Installation
{
"mcpServers": {
"aws-s3": {
"command": "npx",
"args": ["-y", "mcp-server-aws-s3"],
"env": {
"AWS_REGION": "us-east-1",
"AWS_ACCESS_KEY_ID": "AKIA...",
"AWS_SECRET_ACCESS_KEY": "your-secret-key"
}
}
}
}Using AWS Profile
{
"mcpServers": {
"aws-s3": {
"command": "npx",
"args": ["-y", "mcp-server-aws-s3"],
"env": {
"AWS_PROFILE": "mcp-access",
"AWS_REGION": "us-east-1"
}
}
}
}Available S3 Operations
| Operation | Description | Example Prompt |
|---|---|---|
| List buckets | Show all S3 buckets | ”List my S3 buckets” |
| List objects | Objects in a bucket | ”What’s in prod-assets bucket?” |
| Get object | Read object content | ”Read config.json from bucket” |
| Get metadata | Object size, type, date | ”Get metadata for this file” |
EC2 MCP Configuration
Connecting your AI assistant to Amazon EC2 and VPC metadata enables rapid network topology verification, security group ingress auditing, and right-sizing analysis across hundreds of compute instances. Instead of running separate shell commands to cross-reference DescribeInstances, DescribeSecurityGroups, and DescribeSubnets, an agent equipped with the official awslabs.aws-api-mcp-server or mcp-server-aws-ec2 correlates instance ENIs, security group rules, and NAT Gateway routes in a single turn. Because ec2:Describe* actions do not support resource-level permissions in AWS IAM, always enforce aws:RequestedRegion condition keys and tag-based access control (aws:ResourceTag/Environment) if you ever enable start, stop, or reboot actions.
Installation
{
"mcpServers": {
"aws-ec2": {
"command": "npx",
"args": ["-y", "mcp-server-aws-ec2"],
"env": {
"AWS_REGION": "us-east-1",
"AWS_ACCESS_KEY_ID": "AKIA...",
"AWS_SECRET_ACCESS_KEY": "..."
}
}
}
}Available EC2 Operations
| Operation | Description | Example Prompt |
|---|---|---|
| List instances | All EC2 instances | ”Show my EC2 instances” |
| Instance details | Specific instance info | ”Details of i-1234567890” |
| Security groups | List security groups | ”What security groups exist?” |
| VPC info | VPC configuration | ”Show VPC details” |
CloudWatch MCP Configuration
Observability and incident response are among the highest-ROI use cases for AWS MCP servers in production engineering teams. By deploying awslabs.cloudwatch-logs-mcp-server or mcp-server-cloudwatch, site reliability engineers allow Claude or Cursor to execute structured CloudWatch Logs Insights queries (fields @timestamp, @message | filter @message like /ERROR/ | limit 50), correlate metric anomalies across Lambda and RDS, and inspect active composite alarms. Enforcing server-side limit clauses and metric period aggregation (such as 300-second p99 rollups) is critical so a noisy log group does not flood the model context with 50,000+ raw log tokens during an outage.
Installation
{
"mcpServers": {
"cloudwatch": {
"command": "npx",
"args": ["-y", "mcp-server-cloudwatch"],
"env": {
"AWS_REGION": "us-east-1",
"AWS_ACCESS_KEY_ID": "AKIA...",
"AWS_SECRET_ACCESS_KEY": "..."
}
}
}
}Available CloudWatch Operations
| Operation | Description | Example Prompt |
|---|---|---|
| List metrics | Available metrics | ”What metrics are available?” |
| Get metric data | Retrieve metric values | ”CPU usage for last hour” |
| List log groups | CloudWatch log groups | ”Show log groups” |
| Query logs | Search log data | ”Errors in API logs today” |
Practical Workflows
Workflow 1: Infrastructure Inventory
Get a complete view of your AWS resources:
You: Give me an inventory of my AWS infrastructure
Claude: Querying AWS resources...
**AWS Infrastructure Inventory**
**EC2 Instances (12):**
| Name | Type | State | IP |
|---------------|-----------|---------|---------------|
| prod-api-1 | t3.medium | running | 10.0.1.45 |
| prod-api-2 | t3.medium | running | 10.0.1.46 |
| prod-web-1 | t3.large | running | 10.0.2.12 |
| staging-api | t3.small | running | 10.0.3.22 |
| dev-server | t3.micro | stopped | - |
...
**S3 Buckets (8):**
- prod-assets (45 GB)
- prod-backups (320 GB)
- logs-archive (12 GB)
...
**Monthly Estimate:** ~$485Workflow 2: Cost Analysis
Understand your AWS spending:
You: Analyze my S3 storage costs by bucket
Claude: Analyzing S3 storage...
**S3 Cost Analysis**
| Bucket | Size | Storage Class | Monthly Cost |
|------------------|---------|---------------|--------------|
| prod-backups | 320 GB | Standard | $7.36 |
| media-archive | 850 GB | Glacier | $3.40 |
| prod-assets | 45 GB | Standard | $1.04 |
| logs-archive | 12 GB | IA | $0.15 |
| dev-uploads | 2 GB | Standard | $0.05 |
**Total Monthly:** ~$11.95
**Recommendations:**
- Move prod-backups older than 90 days to Glacier (-$4/mo)
- logs-archive could use Glacier Deep Archive (-$0.10/mo)Workflow 3: Security Audit
Review security configurations:
You: Audit my EC2 security groups for overly permissive rules
Claude: Auditing security groups...
**Security Group Audit**
⚠️ **Issues Found (3):**
1. **sg-abc123** (web-servers)
- Port 22 open to 0.0.0.0/0
- Recommendation: Restrict to VPN IP range
2. **sg-def456** (legacy-app)
- Port 3306 open to 0.0.0.0/0
- CRITICAL: MySQL exposed to internet!
- Recommendation: Restrict to app security group
3. **sg-ghi789** (dev-testing)
- All ports open to 0.0.0.0/0
- Recommendation: Delete or restrict this SG
✅ **Compliant (15):**
- All other security groups properly restrictedWorkflow 4: Log Analysis
Search and analyze CloudWatch logs:
You: Find all ERROR logs in the API log group from the last hour
Claude: Querying CloudWatch logs...
**Error Logs (Last Hour)**
Found 23 errors in /aws/lambda/api-handler:
| Time | Error Type | Count |
|----------|---------------------|-------|
| 14:45:32 | TimeoutException | 12 |
| 14:52:18 | ValidationError | 8 |
| 15:01:45 | DatabaseConnection | 3 |
**Top Error:**TimeoutException: Task timed out after 30 seconds at /var/task/handlers/process.js:45 Affected requests: 12
**Recommendation:**
Consider increasing Lambda timeout or optimizing
the database query causing delays.Workflow 5: Resource Monitoring
Real-time resource monitoring:
You: How are my production EC2 instances performing?
Claude: Fetching CloudWatch metrics...
**EC2 Performance Dashboard**
| Instance | CPU | Memory* | Network I/O | Status |
|-------------|-----|---------|-------------|--------|
| prod-api-1 | 45% | 72% | 125 MB/s | ✅ OK |
| prod-api-2 | 52% | 68% | 118 MB/s | ✅ OK |
| prod-web-1 | 23% | 41% | 45 MB/s | ✅ OK |
| prod-db-1 | 78% | 89% | 234 MB/s | ⚠️ High |
**Alerts:**
- prod-db-1: CPU and memory elevated
- Recommendation: Consider scaling or query optimization
*Memory requires CloudWatch agentExample Prompts
S3 Operations
# Bucket listing
"List all my S3 buckets"
"Show buckets in us-west-2"
"Which buckets have public access?"
# Object operations
"List objects in prod-assets bucket"
"Show files uploaded today to uploads/"
"Read the config.json from settings bucket"
# Analysis
"How much storage am I using across all buckets?"
"Which bucket has the most objects?"
"Find large files over 1GB"EC2 Operations
# Instance listing
"Show all running EC2 instances"
"List instances by tag 'Environment=Production'"
"Which instances are in the us-east-1a availability zone?"
# Instance details
"Get details of i-0abc123def456"
"What security groups are attached to my web servers?"
"Show the launch configuration for prod instances"
# Cost and sizing
"Which instances are over-provisioned?"
"Show instance hours for this month"
"List instances that have been stopped for over 30 days"CloudWatch Operations
# Metrics
"Show CPU utilization for the last 24 hours"
"Get memory usage for prod-api instances"
"What's the average latency for my API?"
# Logs
"Search for errors in the last hour"
"Find logs mentioning 'timeout'"
"Show logs from the payment service"
# Alarms
"List active CloudWatch alarms"
"Which alarms are in ALARM state?"
"Show alarm history for this week"Multiple AWS Accounts
Enterprise cloud architectures following the AWS Well-Architected Framework isolate workloads across dedicated organizational units (Production, Staging, Shared Services, and Security Tooling) managed via AWS Organizations and AWS IAM Identity Center. When configuring MCP servers for multi-account operations, never share a single IAM user credential across accounts; instead, register distinct named MCP server instances bound to isolated AWS_PROFILE definitions in ~/.aws/config that assume cross-account roles via sts:AssumeRole with 1-hour session durations. Read-only SecurityAudit or ViewOnlyAccess policies should be enforced on production profiles, while sandbox or development profiles can permit scoped awslabs.cfn-mcp-server or awslabs.terraform-mcp-server validation calls.
For multi-account management:
{
"mcpServers": {
"aws-prod": {
"command": "npx",
"args": ["-y", "mcp-server-aws-s3"],
"env": {
"AWS_PROFILE": "production",
"AWS_REGION": "us-east-1"
}
},
"aws-staging": {
"command": "npx",
"args": ["-y", "mcp-server-aws-s3"],
"env": {
"AWS_PROFILE": "staging",
"AWS_REGION": "us-east-1"
}
},
"aws-dev": {
"command": "npx",
"args": ["-y", "mcp-server-aws-s3"],
"env": {
"AWS_PROFILE": "development",
"AWS_REGION": "us-west-2"
}
}
}
}Then specify which account:
You: Using the production AWS account, list S3 buckets
You: Check EC2 instances in stagingSecurity Best Practices
Credential Management
| ✓ Do | ✗ Don’t |
|---|---|
| Use IAM users/roles | Use root account credentials |
| Apply least-privilege | Grant admin access |
| Rotate keys regularly | Keep old keys active |
| Use AWS profiles | Hardcode in config files |
| Audit access logs | Ignore CloudTrail |
For more on AI security best practices, see the Understanding AI Safety, Ethics, and Limitations guide.
IAM Policy Examples
Minimal S3 Read:
{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": ["s3:GetObject", "s3:ListBucket"],
"Resource": ["arn:aws:s3:::my-bucket", "arn:aws:s3:::my-bucket/*"]
}
]
}EC2 Describe Only:
{
"Effect": "Allow",
"Action": "ec2:Describe*",
"Resource": "*"
}Audit Trail
Enable CloudTrail to track all API calls:
You: Show recent API calls made by the MCP IAM user
Claude: CloudTrail events for mcp-user (last 24h):
- s3:ListBuckets (45 calls)
- s3:GetObject (23 calls)
- ec2:DescribeInstances (12 calls)
No suspicious activity detected.Combining with Other MCPs
The real operational leverage of AWS MCP servers emerges when you compose them with source control, incident alerting, and persistent architectural memory servers inside a unified DevOps agent loop. As detailed in our analysis of AI Agents vs Agentic AI, connecting cloud telemetry directly to code repositories and team communication channels eliminates manual context-switching during Severity-1 incidents. By pairing CloudWatch and EC2 servers with Slack, Filesystem, and Memory MCP servers, an SRE agent can detect a latency regression, correlate it with a recent container deployment, export a compliance report, and brief the on-call channel in under 45 seconds.
AWS + Slack
Alert team about issues:
You: Check for any AWS alarms in ALARM state and notify #ops
Claude: [Uses AWS to check alarms]
[Uses Slack to notify]
Posted to #ops:
🚨 **AWS Alarm Alert**
- HighCPU-ProdAPI: i-abc123 at 95% CPU
- DatabaseConnections: RDS connections maxedAWS + Filesystem
Export AWS data:
You: Export EC2 inventory to a CSV file
Claude: [Uses AWS to query EC2]
[Uses Filesystem to save CSV]
Saved to: /reports/ec2-inventory.csv
- 12 instances documented
- Includes: name, type, state, IPsAWS + Memory
Remember infrastructure patterns:
You: Learn my AWS infrastructure and remember it
Claude: [Uses AWS to analyze resources]
[Uses Memory to store patterns]
Memorized:
- 3 environments: prod, staging, dev
- Naming convention: env-service-number
- Regions: us-east-1 (prod), us-west-2 (dev)Troubleshooting
Diagnosing AWS MCP server failures typically involves inspecting three layers: local runtime dependencies (uvx for Python-based awslabs servers or npx for Node.js servers), AWS STS credential resolution, and IAM policy evaluation logic. When an AWS API call fails, AWS returns a structured error code (AccessDeniedException, ExpiredTokenException, or ThrottlingException) that you can trace in AWS CloudTrail within 60 to 180 seconds of invocation. If a policy blocks an expected read call, run the IAM Policy Simulator or check whether a Service Control Policy (SCP) or VPC Endpoint Policy is restricting access from your workstation network.
Issue: “Access Denied”
Symptoms: Permission errors
Solutions:
| Cause | Solution |
|---|---|
| Missing IAM permissions | Add required policies |
| Wrong credentials | Check keys are correct |
| Wrong region | Verify region in config |
| Resource policy blocking | Check bucket/resource policies |
Issue: “Credential not found”
Symptoms: Authentication fails
Solutions:
- Check env variables are set
- Verify AWS profile exists
- Ensure credentials aren’t expired
- Check AWS STS for temporary creds
Issue: “Rate limited”
Symptoms: Too many requests
Solutions:
- Add delays between operations
- Reduce query scope
- Check AWS service limits
Related MCP Servers
| Server | Complements AWS By… |
|---|---|
| Cloudflare MCP | Edge DNS, Workers, and R2 integration |
| Slack MCP | Alerting and notifications |
| Filesystem MCP | Saving AWS data locally |
| Memory MCP | Remembering infrastructure |
| PostgreSQL MCP | Correlating with app data |
Summary
AWS MCP Servers bring AI to cloud infrastructure:
- ✓ S3 - Bucket and object management
- ✓ EC2 - Instance monitoring and details
- ✓ CloudWatch - Metrics and logs
- ✓ Multi-account - Manage multiple environments
- ✓ Secure - IAM-based access control
- ✓ Official AWS (2025) - EKS, ECS, Lambda MCP servers
2025 AWS MCP Ecosystem:
- Amazon EKS MCP - Kubernetes cluster management
- Amazon ECS MCP - Container service orchestration
- AWS Lambda MCP - Serverless function integration
- AWS API MCP - Natural language AWS API access
- Amazon Q Developer - AI-powered infrastructure
Best use cases:
- Infrastructure inventory
- Cost analysis
- Security audits
- Log analysis
- Resource monitoring
Security checklist:
- ☐ Create dedicated IAM user for MCP
- ☐ Apply least-privilege policies
- ☐ Enable CloudTrail logging
- ☐ Rotate credentials regularly
- ☐ Never use root credentials
Next: Learn about Cloudflare MCP Server → for edge infrastructure.
By pairing least-privilege IAM roles with the official AWS Labs servers, your engineering team can safely scale S3 AI automation and cloud automation AI across staging and production.
Questions about AWS MCP? Check the Official AWS Labs MCP Repository on GitHub, the AWS MCP Documentation Portal, or the Model Context Protocol Specification.