Mara-Terra
MCP server exposing internal Terraform and pipeline module knowledge to AI assistants through Azure AI Search.
Overview
Challenge
Engineers had hundreds of internal Terraform and pipeline modules spread across sprawling repos and documentation. Finding the right module meant asking someone who already knew — tribal knowledge as a dependency.
Solution
Built an MCP (Model Context Protocol) server from scratch in Python and FastMCP, backed by Azure AI Search indices for Terraform modules, pipeline modules, and documentation. Domain-specific providers expose read-only, idempotent tool contracts over both stdio and HTTP, so any MCP-compatible AI client can query internal infrastructure knowledge directly. Dual-credential identity — Azure CLI locally, managed identity in cloud — with OData sanitization, bounded results, and retry logic.
Impact
Turned undocumented infrastructure knowledge into a searchable surface that AI assistants can reach, removing the manual hunt through module repos and the dependency on tribal knowledge.
Tech Stack
Key Metrics
- 7 read-only tools exposed to any MCP-compatible AI client
- Hundreds of Terraform and pipeline modules made searchable
- Built and shipped from scratch in 2 months
Interested in discussing this project or similar work?
Get in Touch