Building a Local MCP Server to Automate Daily DevOps Tasks
I got tired of switching between terminals, running the same commands, and checking the same things every day.
So I built something simple: a local MCP (Model Context Protocol) server that understands what I'm asking — and just does it. No cloud. No API keys. Runs entirely on my laptop using FastAPI and Phi-3 Mini via GPT4All.
🧠 MCP — Three Core Primitives
MCP (Model Context Protocol) works on three building blocks:
🔧 Tools — actions it can execute
- "start the frontend"
- "restart a web app"
- "check ML endpoint logs"
📦 Resources — context it can read
Current environment state, service status, config values — all available to the agent as structured context.
📝 Prompts — reusable workflows
Define a sequence once, trigger it by name. "morning check" automatically runs a chain of tools — no manual steps.
You define what's exposed. The agent just speaks your language and acts on it.
⚙️ Under the Hood
| Component | Tech |
|---|---|
| MCP Server | FastAPI |
| Local LLM | Phi-3 Mini via GPT4All |
| Routing | Hybrid (keyword → fast, LLM → when needed) |
| Tool definitions | Modular JSON |
🔐 Fully Local
Prompts, infrastructure details, internal workflows — nothing leaves my machine. No SaaS subscriptions, no data residency concerns.
💡 Why I Built It
Not for AI hype — just to move faster and stop doing repetitive work.
It feels like a lightweight personal assistant for daily cloud tasks, but fully under my control. The hybrid routing is the key — keywords handle the 80% of fast, predictable tasks instantly, and the LLM only kicks in when you need flexible natural language interpretation.
What would you automate first with something like this?
Related Topics
If you're exploring this space, these are worth researching alongside MCP:
- LangChain / LangGraph — for orchestrating multi-step agent workflows
- Ollama — alternative local LLM runtime (llama3, mistral, phi-3)
- OpenAI function calling — cloud equivalent of MCP tools
- AutoGen / CrewAI — multi-agent frameworks that pair well with MCP servers
- FastAPI + Pydantic — the backbone for building typed, validated MCP tool endpoints