Project 3 — ToolBridge: an MCP Server deployed to the cloud

A minimal Model Context Protocol (MCP) server exposing two tools (get_weather, list_files) plus an HTTP FastAPI wrapper for easy demos. Dockerfile included for containerization; deployed live to Azure App Service on the free tier.

Live demo: https://toolbridge-bhavana-2026.azurewebsites.net/docs

What is MCP?

MCP (Model Context Protocol) is an open standard from Anthropic for connecting LLMs to tools, data sources, and prompts through a common protocol — like USB-C for AI. Any MCP-compatible client (Claude, Cursor, your own LangChain agent) can talk to any MCP server.

Why enterprises care: instead of writing a bespoke integration for every AI assistant, expose internal systems once as an MCP server — authentication, tool schemas, and versioning all handled via the protocol.

Tools exposed

Tool	Input	Returns
get_weather	{city: str}	mock weather payload
list_files	{directory: str}	list of files in an allowlisted directory

Tech stack

• mcp SDK (Python)
• FastAPI (HTTP wrapper for demos)
• Docker for packaging
• Azure App Service / AWS App Runner for hosting

Run the MCP server locally (Windows)

py -3.11 -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python mcp_server.py            :: runs silently over stdio — Ctrl+C to stop

Test with MCP Inspector (the official Anthropic tool)

npx @modelcontextprotocol/inspector python mcp_server.py

Opens an interactive UI at http://localhost:6274 where you can list, call, and inspect responses from your tools.

Run the HTTP demo locally

uvicorn http_bridge:app --reload

Open http://localhost:8000/docs. Try:

• POST /tools/get_weather with {"city": "Mumbai"}
• POST /tools/list_files with {"directory": "."}
• POST /tools/list_files with {"directory": "C:\\Windows"} → returns 403 (allowlist guard)

Deploy to Azure App Service (Free tier)

The exact commands used to ship this to a public URL:

az login
az webapp up --runtime "PYTHON:3.11" --name <your-unique-name> --sku F1 --location "centralindia" --resource-group toolbridge-rg
az webapp config set --resource-group toolbridge-rg --name <your-unique-name> --startup-file "uvicorn http_bridge:app --host 0.0.0.0 --port 8000"
az webapp restart --resource-group toolbridge-rg --name <your-unique-name>

Your app will be live at https://<your-unique-name>.azurewebsites.net/docs.

Why this stack

• MCP — standardizes how AI clients discover + call tools, avoiding bespoke integrations
• FastAPI HTTP wrapper — same tools exposed as REST for non-MCP clients (demos, Postman, Swagger)
• Docker — reproducible packaging; Dockerfile included for anyone wanting to run in a container
• Azure App Service F1 — free-tier PaaS with native Python runtime, no OS patching, auto-SSL

What I learned

• MCP's transport modes (stdio for local, SSE for remote) and when to use each
• Tool schema design — JSON Schema inputSchema matters for how clients discover parameters
• Implementing an allowlist security guard so the AI can't arbitrarily read the filesystem
• The practical difference between Azure's zip-deploy (native Python runtime) and container-deploy (Docker image)
• Why "native services independent of VM" matters — no OS patching, auto-scaling, managed HTTPS out of the box