tools-only ADR-002: Azure Pricing MCP Server Integration
Status
Accepted
Date
2024-01-15
Context
Cost estimation is a critical part of infrastructure planning. Cloud architects and IT professionals need accurate cost information to:
- Create budgets and business cases
- Compare architectural alternatives
- Optimize resource selection (SKUs, regions)
- Provide stakeholders with expected monthly costs
The Problem
GitHub Copilot has training data that may be outdated for Azure pricing. Azure prices change frequently due to:
- New SKU introductions
- Regional pricing variations
- Promotional offers and reserved capacity discounts
- Currency fluctuations
Relying on Copilot's built-in knowledge would result in inaccurate cost estimates.
Considered Alternatives
- Use Azure Pricing Calculator links - Direct users to the web calculator
- Embed static pricing data - Include pricing tables in agent prompts
- Call Azure Retail Prices API directly - Use terminal commands to fetch prices
- Build custom MCP server - Dedicated pricing service with caching
Decision
We built a custom Model Context Protocol (MCP) server for Azure pricing that:
- Integrates with the Azure Retail Prices API
- Provides 6 specialized tools for different pricing scenarios
- Caches responses for 1 hour to reduce API calls
- Uses singleton HTTP session for connection reuse
- Includes customer discount handling (disabled by default)
MCP Server Tools
| Tool | Purpose |
|---|---|
azure_price_search |
Search prices with filters (service, SKU, region) |
azure_price_compare |
Compare prices across regions or SKUs |
azure_region_recommend |
Find cheapest region for a service |
azure_cost_estimate |
Estimate monthly costs based on usage |
azure_discover_skus |
List available SKUs for a service |
get_customer_discount |
Get/set customer discount percentage |
Why MCP?
- Native Copilot integration - Tools appear in Copilot's tool list
- Real-time data - Always fetches current Azure prices
- Structured responses - Returns formatted data Copilot can reason about
- Extensible - Easy to add new pricing tools
Consequences
Positive
- Cost estimates use real-time Azure pricing data
- Agents can compare regions and recommend cost-optimized options
- Reduces risk of outdated pricing in generated documentation
- Demonstrates MCP extensibility for enterprise scenarios
- Caching reduces API latency for repeated queries
Negative
- Requires Python environment and dependencies
- Dev container setup needed for seamless experience
- API calls add latency to responses (~2-5 seconds)
- Pricing API has rate limits (may affect heavy usage)
Mitigations
- Pre-configured in dev container with automatic setup
- 1-hour TTL cache reduces API calls
- Connection pooling for efficient HTTP handling
- Graceful error handling with fallback messages
Pricing Data Accuracy
Important: The MCP server uses the Azure Retail Prices API which provides:
- ✅ Real-time list prices (pay-as-you-go rates)
- ✅ Current SKU availability by region
- ❌ Does NOT include negotiated EA/CSP discounts
- ❌ Does NOT include reserved instance pricing by default
- ❌ Does NOT include promotional offers
All cost estimates should be validated with the Azure Pricing Calculator for production budgeting.
Implementation Details
Directory Structure
mcp/azure-pricing-mcp/
├── src/azure_pricing_mcp/
│ ├── server.py # MCP server with caching, session management
│ └── handlers.py # Tool implementations
├── requirements.txt # Dependencies (aiohttp, cachetools, mcp)
└── README.md # Setup and usage documentation
Key Technical Decisions
- Singleton HTTP session - Prevents "Connector is closed" errors
- TTLCache (1 hour) - Balances freshness vs. performance
- 30-second timeout - Prevents hanging on slow API responses
- Discount default = 0% - User must explicitly request discounts
References
- mcp/azure-pricing-mcp/README.md - Server documentation
- Azure Retail Prices API
- Model Context Protocol - MCP specification