MCP CloudOps Demo

A free, public, portfolio-friendly Model Context Protocol (MCP) demo that shows how an AI client can discover and use CloudOps capabilities without requiring a real Azure subscription, API key, or production credentials.

The server exposes tools, resources, and prompts backed by deterministic fake Azure-style infrastructure data. It is designed to be easy to clone, run locally, inspect with MCP Inspector, containerize, and later extend with real Azure or GitHub integrations.

Safety: this repository never connects to real infrastructure by default. Even the restart capability is a simulation.

What this demonstrates

A normal chatbot only knows what is in its conversation context. An MCP-enabled client can also discover structured capabilities exposed by an MCP server.

In this demo, the client can:

• discover production and staging servers;

• inspect server health and utilization;

• correlate incidents, deployments, and logs;

• read infrastructure inventory as MCP resources;

• use reusable incident-investigation prompts;

• request a simulated operational action with explicit confirmation.

Related MCP server: mcp-azure-toolkit

Architecture

flowchart LR
    U[User] --> C[MCP-compatible AI client]
    C <-->|MCP over stdio| S[MCP CloudOps Demo Server]

    S --> T[Tools]
    S --> R[Resources]
    S --> P[Prompts]

    T --> D[(Fake Azure-style JSON data)]
    R --> D
    P --> C

    T --> T1[list_servers]
    T --> T2[get_server_health]
    T --> T3[get_recent_deployments]
    T --> T4[get_open_incidents]
    T --> T5[search_logs]
    T --> T6[restart_demo_service]

A larger diagram is also available in docs/architecture.md.

MCP primitives used

Tools

Resources

Prompts

Demo scenario

Suppose the user asks:

Which production servers are unhealthy or under resource pressure?

An MCP client can discover and call list_servers(environment="prod"), see that api-prod-02 is degraded with high CPU, and then decide to call get_server_health or inspect incidents and logs.

A follow-up might be:

Investigate the issue affecting payments-api and tell me whether it is related to a deployment.

The client can combine multiple MCP capabilities:

User request
   |
   v
get_open_incidents("payments-api")
   |
   v
list_servers("prod")
   |
   v
get_server_health("api-prod-02")
   |
   v
get_recent_deployments(service="payments-api")
   |
   v
search_logs(service="payments-api")
   |
   v
AI correlates evidence and explains the likely cause

This is the core value proposition of MCP: the model can work with standardized external capabilities instead of having every integration hard-coded into the chat application.

Repository structure

mcp-cloudops-demo/
├── .github/
│   └── workflows/
│       └── ci.yml
├── docs/
│   └── architecture.md
├── examples/
│   └── prompts.md
├── src/
│   └── mcp_cloudops/
│       ├── data/
│       │   ├── deployments.json
│       │   ├── incidents.json
│       │   ├── logs.json
│       │   └── servers.json
│       ├── prompts/
│       │   └── cloudops.py
│       ├── resources/
│       │   └── cloudops.py
│       ├── tools/
│       │   └── cloudops.py
│       ├── web/
│       │   ├── app.py
│       │   ├── scenario.py
│       │   └── static/
│       │       ├── index.html
│       │       ├── styles.css
│       │       └── app.js
│       ├── server.py
│       └── store.py
├── tests/
│   ├── test_data_relationships.py
│   └── test_store.py
├── Dockerfile
├── docker-compose.yml
├── LICENSE
├── pyproject.toml
└── README.md

Requirements

• Python 3.11 or newer

• pip, uv, or another Python package manager

• Optional: Docker

• Optional: an MCP-compatible client or MCP Inspector

Quick start with Python

Clone the repository and create a virtual environment:

git clone https://github.com/BozyBonifacio/mcp-cloudops-demo.git
cd mcp-cloudops-demo
python -m venv .venv

Activate it.

Linux/macOS:

source .venv/bin/activate

Windows PowerShell:

.\.venv\Scripts\Activate.ps1

Install the project:

python -m pip install -e '.[dev]'

Run the MCP server over stdio:

mcp-cloudops-demo

The process will wait for an MCP client to communicate over stdin/stdout. That is expected.

Detailed Windows run guide

From PowerShell:

git clone https://github.com/BozyBonifacio/mcp-cloudops-demo.git
cd mcp-cloudops-demo
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"
mcp-cloudops-demo

If PowerShell blocks virtual-environment activation, allow scripts for the current terminal session only:

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass
.\.venv\Scripts\Activate.ps1

When mcp-cloudops-demo starts and appears to sit idle, that is expected. This demo uses MCP over stdio and waits for an MCP client rather than opening a web page. Press Ctrl+C to stop it.

Browser-based portfolio demo

The easiest way to demonstrate the project is now the built-in browser UI. It requires no paid LLM API key and acts as a real MCP client: the FastAPI backend starts the local MCP server over stdio, discovers its capabilities, calls MCP tools, and returns the evidence to the browser.

Start it after installing the project:

mcp-cloudops-web

Then open:

http://localhost:8000

The page includes:

• an operator-style chat interface;

• suggested CloudOps investigation prompts;

• runtime MCP capability discovery;

• an MCP trace showing each tool call and its arguments;

• a deterministic incident-investigation workflow that requires no external AI service;

• the existing safety behavior for the simulated restart.

A recommended first prompt is:

Investigate the open incident affecting payments-api and tell me the likely cause.

The browser backend will use MCP to call multiple tools such as get_open_incidents, list_servers, get_recent_deployments, and search_logs, then display both the conclusion and the MCP call trace.

Why deterministic chat? The browser demo intentionally does not require OpenAI, Azure OpenAI, Anthropic, or another paid model API. A small intent router recognizes the included demo scenarios while MCP remains responsible for capability discovery and tool execution. You can later replace the router with an LLM without changing the MCP server contract.

Browser demo architecture

Browser
   |
   | HTTP /api/chat
   v
FastAPI browser backend
   |
   | MCP over stdio
   v
MCP CloudOps Server
   |
   +--> Tools
   +--> Resources
   +--> Prompts
   |
   v
Fake Azure-style JSON data

Browser demo with Docker

Build and start the container:

docker compose up --build

Then open:

http://localhost:8000

Stop it with Ctrl+C, then optionally remove the container with:

docker compose down

Test with MCP Inspector

The official Python MCP SDK includes development tooling when installed with the CLI extra.

If your environment has the MCP CLI available, run:

mcp dev src/mcp_cloudops/server.py

Then open the Inspector URL printed in your terminal. Explore the Tools, Resources, and Prompts tabs.

If mcp is not installed as a CLI command, install the SDK CLI extra:

python -m pip install 'mcp[cli]'

Recommended MCP Inspector walkthrough

After starting Inspector, use this sequence to demonstrate how multiple MCP capabilities contribute to one investigation:

1. Call list_servers with environment = prod.

2. Call get_server_health with server_name = api-prod-02.

3. Call get_open_incidents and investigate payments-api.

4. Call get_recent_deployments for payments-api.

5. Call search_logs for the affected service/server and inspect ERROR events.

list_servers("prod")
        |
        v
get_server_health("api-prod-02")
        |
        v
get_open_incidents("payments-api")
        |
        v
get_recent_deployments(service="payments-api")
        |
        v
search_logs(service="payments-api")
        |
        v
AI correlates the evidence

Recommended live-demo prompt

Investigate the open incident affecting payments-api. Determine which server is affected, its health state, whether a recent deployment correlates with the incident, what the logs show, and the most likely cause.

The key point to explain is that the MCP server advertises standardized capabilities that an MCP-compatible host can discover and combine instead of relying on one hard-coded chatbot workflow.

Windows MCP client configuration

When using a virtual environment on Windows, prefer the absolute path to its Python executable. For example, if the repository is at C:\github\mcp-cloudops-demo:

{
  "mcpServers": {
    "cloudops-demo": {
      "command": "C:\\github\\mcp-cloudops-demo\\.venv\\Scripts\\python.exe",
      "args": ["-m", "mcp_cloudops.server"]
    }
  }
}

Replace the example path with your actual clone location. This avoids accidentally launching a different Python installation that does not contain the project dependencies.

Troubleshooting

If mcp-cloudops-demo is not recognized, activate .venv and reinstall the project:

.\.venv\Scripts\Activate.ps1
python -m pip install -e ".[dev]"

You can also try:

python -m mcp_cloudops.server

If mcp is not recognized:

python -m pip install "mcp[cli]"
mcp dev src/mcp_cloudops/server.py

If you get ModuleNotFoundError: mcp_cloudops, make sure you are in the repository root, reinstall the editable package, and verify the import:

python -m pip install -e ".[dev]"
python -c "import mcp_cloudops; print('mcp_cloudops import OK')"

Run tests

pytest

Run linting:

ruff check src tests

Run with Docker

The Docker image now starts the browser UI by default.

docker build -t mcp-cloudops-demo .
docker run --rm -p 8000:8000 mcp-cloudops-demo

Open http://localhost:8000.

Or use Docker Compose:

docker compose up --build

To run the original stdio MCP server inside the image instead:

docker run --rm -i --entrypoint mcp-cloudops-demo mcp-cloudops-demo

Example client configuration

MCP hosts generally need only the command that starts the stdio server. The exact configuration format depends on your client.

A typical local configuration conceptually looks like this:

{
  "mcpServers": {
    "cloudops-demo": {
      "command": "python",
      "args": ["-m", "mcp_cloudops.server"]
    }
  }
}

If the package is installed into a virtual environment, point the host at that environment's Python executable or use the installed mcp-cloudops-demo command.

Suggested prompts

Try:

• Which production servers are unhealthy or under resource pressure?

• Investigate the open incident affecting payments-api.

• Did a recent deployment correlate with the current production issue?

• Show ERROR logs for api-prod-02 and explain the likely cause.

• Use the investigate_incident prompt for payments-api.

• Read infra://inventory/production and summarize capacity risks.

• Simulate restarting api-prod-02, but ask me for confirmation first.

More examples are in examples/prompts.md.

Why fake data?

A public MCP demo should be runnable by anyone without exposing secrets or requiring cloud spend. The JSON fixtures provide:

• zero infrastructure cost;

• deterministic behavior for demonstrations;

• safe public-source control;

• repeatable automated tests;

• a clean boundary between MCP capabilities and the eventual real data source.

Extending the demo to real Azure

The easiest upgrade path is to keep the MCP-facing functions and replace load_json() with an adapter layer.

For example:

MCP Tool
   |
   v
CloudOps service interface
   |
   +--> DemoJsonProvider
   |
   +--> AzureProvider
          |
          +--> Azure Resource Graph
          +--> Azure Monitor
          +--> Log Analytics
          +--> Azure DevOps / GitHub

Potential real integrations include:

• Azure Resource Graph for VM/resource inventory;

• Azure Monitor for metrics;

• Log Analytics for KQL queries;

• Azure Update Manager for patch posture;

• GitHub for commits, pull requests, and Actions runs;

• Azure DevOps for pipelines and deployment history.

For a public portfolio version, prefer read-only permissions and use environment variables or managed identity rather than storing credentials in the repository.

Example future multi-server scenario

A stronger second version of this project could demonstrate an AI client correlating data across multiple MCP servers:

Developer
  |
  | "Why did the deployment fail?"
  v
AI / MCP Host
  |
  +--> GitHub MCP server ------> commit / pull request
  |
  +--> Azure MCP server -------> deployment status
  |
  +--> Observability MCP ------> application logs
  |
  v
Correlated incident explanation

That demonstrates why a protocol is useful: each domain can expose capabilities independently while the host provides the conversational orchestration.

Security considerations

This repository intentionally follows several safe-demo patterns:

• no secrets are committed;

• no cloud credentials are required;

• operational data is fictional;

• write-like actions are simulations;

• the simulated restart requires an explicit confirmed=true parameter;

• GitHub Actions uses read-only repository contents permission;

• tests verify fixture relationships so demos stay coherent.

If you replace the fake provider with real infrastructure, add authentication, authorization, audit logging, least-privilege access, input validation, and confirmation controls before exposing mutating tools.

CI

The included GitHub Actions workflow runs on pushes and pull requests and validates:

• Python 3.11;

• Python 3.12;

• Ruff linting;

• Pytest tests.

Useful MCP references

• Model Context Protocol: https://modelcontextprotocol.io/

• Official Python SDK: https://github.com/modelcontextprotocol/python-sdk

• MCP specification: https://github.com/modelcontextprotocol/modelcontextprotocol

License

MIT. See LICENSE.

Portfolio talking points

When demonstrating this repository, emphasize these points:

1. The AI host is separate from the MCP server.

2. The server advertises capabilities instead of embedding chatbot logic.

3. Tools are actions/queries, resources are addressable context, and prompts are reusable workflows.

4. The same MCP server can be consumed by different compatible hosts.

5. The fake-data provider can be replaced without redesigning the MCP interface.

6. Mutating operations should have stronger authorization and confirmation controls than read operations.

Built as an educational CloudOps example. All infrastructure names, incidents, deployments, and metrics in the default dataset are fictional.