DevOps-MCP-Server

A Model Context Protocol (MCP) server that centralizes an organization's skills, instructions, and policies for AI coding agents, so tools like Claude, ChatGPT, and Gemini generate output that follows organizational standards from the first prompt, rather than requiring iterative correction.

I researched, designed, and built this entirely on my own: the MCP server itself, the structured skills and policy content it serves, and the centralized, authenticated hosting model that lets an entire organization stay in sync from a single source rather than distributing files by hand.

Getting consistently accurate, standards-compliant output from an AI coding agent like ChatGPT, Claude, or Gemini usually takes more than one prompt. Ask an agent to build a Terraform project to deploy a Container Apps environment, for example, and without specific guidance it will structure the project however it sees fit: how variables and models are defined, where secrets are stored, what logging is included, and how the project is organized are all left to the agent's own defaults, often requiring several follow-up prompts to correct before the output matches organizational conventions.

Providing an AI agent with skill and instruction files on a local device solves this for one person, but it doesn't scale. Across an organization with many people using AI tools, there's no practical way to distribute those files to everyone and keep them in sync: every time a standard or policy changes, the updated files have to be manually redistributed, with no guarantee everyone is actually working from the latest version.

1. Researched how to centralize AI agent skills, instructions, and policies so they could be maintained in one place and consumed by any AI tool, rather than distributed as local files per device.

2. Designed the content model first: versioned Markdown files with YAML frontmatter, organized by category (Python, Ansible, Terraform, Docker, Kubernetes, CI/CD, and general engineering practices) and type (skills, instructions, prompts, policies, and rules), so new organizational standards can be added as content rather than code changes.

3. Built the MCP server itself using FastMCP mounted on a Starlette ASGI application, exposing four tools for searching, listing, and retrieving that content, so a connected AI agent can look up exactly what it needs, for instance how to define infrastructure variables, where secrets should be stored, or what logging a project needs.

4. Deliberately restricted every exposed tool to read-only operations, so AI agents get a controlled, auditable window into these standards with no ability to modify them.

5. Secured the endpoint with API key authentication and containerized the server with Docker, so it can be deployed centrally behind auth rather than run locally per device.

6. Designed the setup so each user configures their AI tool to point at this one MCP server endpoint only once; from that point on, every prompt automatically has access to whatever the current organizational standards are, with no further per-user setup.

7. Built this so that updating a policy or skill only means updating its Markdown file and restarting the server; every connected AI agent picks up the change immediately, with nothing to redistribute.

8. Published the project as open source on GitHub, where it was subsequently auto-indexed by mcpmarket.com, making it discoverable to any MCP-compatible AI assistant.