Designing an AI-Native FinOps Layer on AWS with Official MCP Servers (No Custom Code Needed)

What This Guide Covers

This is a step-by-step architecture and implementation blog to build an AI-powered FinOps system on AWS using official or supported MCP servers, without writing custom boto3 code or scripts.

We’ll walk through:

Choosing and deploying the right MCP servers
Wiring them into a single intelligent layer
Letting AI agents query cost, usage, pricing, and optimization data
Real prompts and architecture patterns

Outcome: You can ask Claude or Amazon Q:
“Which services had the highest cost increase over the last quarter?”
“What’s the hourly price of r6g.large in ap-south-1 right now?”
“Do I have underutilized EC2 reservations?”

And get back live, structured responses from your real AWS accounts.

The Stack (No Custom Code Required)

Layer	Tools/Services Used
AI Agent	Claude, Amazon Q, Cursor
MCP Connectors	AWS Pricing MCP

MCP Servers You Will Use

1. AWS Pricing MCP Server

Purpose: Live price queries for any AWS service/SKU/region
Repo: github.com/awslabs/aws-pricing-mcp-server
Example Prompt:

“What is the hourly rate of m5.4xlarge in eu-central-1 with Linux on-demand?”

2. AWS API MCP Server

Purpose: Exposes AWS service documentation, CLI syntax, parameter structure
Use case: Helps LLM reason through what APIs exist and how they can be used
Example Prompt:

“How do I list EC2 volumes and their sizes using the AWS CLI?”

3. Cost Explorer MCP Server

Purpose: Surface high-level usage and cost insights from Cost Explorer
Features:

Filter by service, time, account, tag
Summarize costs by dimension
Built-in natural language routing

Example Prompt:

“Show me EC2 cost trend for ap-south-1 over the last 3 months”
“Which account had the highest spend growth between Q1 and Q2?”

Deploying MCP Servers

Step 1: Clone and Run the Official MCP Servers

All servers are containerized and work with Claude/Amazon Q clients.

Run Pricing MCP:

🐳 Run AWS API MCP:

Run Cost Explorer MCP:

🛡 For security: mount an AWS profile or use environment variables with read-only permissions.

IAM Role Setup

You need an IAM role or user that allows:

Use this in an AWS profile like ai-finops-role and configure each MCP container to use this.

How the AI Workflow Works (No Boto3)

User Prompt → Agent → MCP Server → Result

Here’s what actually happens:

Prompt	Routed To	Returned
“How much did Lambda cost in June in us-west-2?”	Cost Explorer MCP	JSON summary by service, region
“What is the on-demand price of c7g.xlarge in Mumbai?”	Pricing MCP	Live price w/ region, offering class
“Do I have unused EC2 RIs?”	Cost Explorer MCP	RI utilization metrics
“How do I list S3 buckets?”	AWS API MCP	CLI syntax & docs

No glue code. No API wrappers. All managed through the MCP interfaces.

Real Prompt Examples

Use Case	Prompt
Cost trend	“Show EC2 spend trend over last 6 months in ap-south-1”
Tag filter	“Give me costs for resources with tag `project:alpha` in May”
Optimization	“Which account had low RI utilization last month?”
Forecast	“What’s the projected spend for July 2025?”
Live pricing	“Current spot and on-demand prices for r6i.large in us-east-1”
API syntax	“How to call GetCostAndUsage from AWS CLI?”

🛡 Security Notes

No write permissions needed
Every MCP server logs requests to local console for audit
Optional: Run servers inside secure ECS cluster with endpoint routing

🧩 Optional: Frontend or CLI Integration

You can plug your MCP stack into:

Claude Desktop or Claude API Agent
Amazon Q Developer CLI
Internal VS Code Extension
Streamlit-based frontend for FinOps stakeholders

If you want, I can help build these on top—zero boto3 required.

✅ Summary

What You Get	How
Live AWS pricing insights	Pricing MCP
Cost trends, breakdowns	Cost Explorer MCP
API reasoning, CLI help	AWS API MCP
No code, no wrappers	Fully MCP-based
AI-ready prompts	Via Claude / Q / any LLM with MCP plugin support