Cloud FinOps Analyst — MCP Server

An AI-powered AWS cost analysis tool built as a Model Context Protocol (MCP) server. Connect it to Claude and ask plain-English questions about your cloud spend.

What it does

Instead of digging through AWS Cost Explorer dashboards, just ask:

• "What did we spend last month and where is the money going?"
• "Did we have any unexpected cost spikes this week?"
• "Which resources are we paying for but not using?"
• "Which instances are over-provisioned and how much could we save?"
• "How bad is our tagging compliance and how much spend is unallocated?"

Claude calls the right tools, pulls the data, and gives you a clear analysis with specific recommendations.

---

Tools

Tool	Description
get_cost_summary	Cost breakdown by service and region for last 7 days, 30 days, or 3 months
detect_cost_anomalies	Flags unusual spending spikes with likely root causes
get_idle_resources	Finds unused EC2 instances, RDS databases, and unattached EBS volumes
get_rightsizing_recommendations	Suggests downsizing over-provisioned instances with estimated savings
get_tagging_compliance	Identifies untagged resources and unallocatable spend

---

Architecture

You (chat in Claude Desktop or Claude.ai)
        ↓
Claude (AI reasoning — decides which tools to call)
        ↓
This MCP Server (fetches and returns data)
        ↓
AWS APIs (Cost Explorer, EC2, RDS, S3)

---

Getting Started

Prerequisites

• Node.js 18+
• Claude Desktop (free) — download here

Installation

git clone https://github.com/yourusername/finops-mcp-server
cd finops-mcp-server
npm install
npm run build

Connect to Claude Desktop

Add this to your Claude Desktop config file:

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

{
  "mcpServers": {
    "finops-analyst": {
      "command": "node",
      "args": ["/absolute/path/to/finops-mcp-server/dist/index.js"]
    }
  }
}

Restart Claude Desktop. You'll see the tools icon appear in the chat interface.

---

AWS Setup (for real data)

The server currently runs with mock data. To connect to a real AWS account:

1. Create an IAM user with read-only permissions:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "ce:GetCostAndUsage",
        "ce:GetCostForecast",
        "ce:GetAnomalies",
        "ec2:DescribeInstances",
        "ec2:DescribeVolumes",
        "rds:DescribeDBInstances",
        "cloudwatch:GetMetricStatistics",
        "tag:GetResources"
      ],
      "Resource": "*"
    }
  ]
}

2. Add credentials to your environment:

export AWS_ACCESS_KEY_ID=your_key
export AWS_SECRET_ACCESS_KEY=your_secret
export AWS_DEFAULT_REGION=us-east-1

3. Replace the mock data calls in each tool file with real AWS SDK calls.

---

Example Conversations

Cost overview:

"Give me a breakdown of our AWS spend last month"

Anomaly investigation:

"Why did our bill spike last week? What caused it?"

Cost optimization:

"What are the quickest wins to reduce our AWS bill right now?"

Executive report:

"Summarize our cloud cost health and give me 3 priority actions I can bring to my manager"

---

Project Structure

finops-mcp-server/
├── src/
│   ├── index.ts              # MCP server entry point + tool registry
│   ├── tools/
│   │   ├── costSummary.ts        # get_cost_summary
│   │   ├── anomalyDetection.ts   # detect_cost_anomalies
│   │   ├── idleResources.ts      # get_idle_resources
│   │   ├── rightsizing.ts        # get_rightsizing_recommendations
│   │   └── taggingCompliance.ts  # get_tagging_compliance
│   └── data/
│       └── mockData.ts           # Realistic mock AWS data for demos
├── dist/                     # Compiled JavaScript (generated)
├── package.json
├── tsconfig.json
└── README.md

---

Roadmap

• [ ] Connect to real AWS Cost Explorer API
• [ ] Add Reserved Instance vs On-Demand comparison tool
• [ ] Add savings plan coverage analysis
• [ ] Phase 2: Web app with React frontend + Express backend

---

Tech Stack

• TypeScript — type-safe tool definitions
• MCP SDK (@modelcontextprotocol/sdk) — server protocol
• Zod — input schema validation
• Node.js — runtime