Costa Rica
Last updated: 2025-11-03
MCP is about
structured behavior, access control, and responsibilitiesfrom the AI’s perspective, and we expose it (often via HTTP) using whatever hosting option fits best.
Tip
You can think of MCP as:
- A universal API contract for AI agents.
- A permissions framework (AI can only do what’s declared).
- A deployment‑agnostic service (you choose where/how to host it).
MCP (Model Context Protocol) is a structured contract between an AI client (like Copilot Studio or Azure AI Foundry) and an external service (your MCP server).
It defines:
- What tools exist (functions the AI can call).
- What inputs they require (schemas).
- What outputs they return (structured JSON).
- What resources are available (read‑only context like docs, schemas, or files).
- What prompts are predefined (templates the AI can use).
Tip
Like a set of rules and responsibilities that tell the AI: "Here’s what you’re allowed to do, here’s how you call it, and here’s what you’ll get back"
From the AI’s perspective:
- Rights: It can only call the tools/resources the MCP server advertises.
- Responsibilities: It must respect the input/output schema and handle errors gracefully.
- Boundaries: The AI cannot
“invent”new tools, it only uses what the MCP server exposes.
From developer perspective (as the server owner):
- You decide what to expose (e.g.,
getCustomerOrders,createInvoice). - You enforce security and governance (auth, rate limits, logging).
- You control where it’s hosted (local dev, Azure App Service, Container Apps, Functions, etc.).
- MCP itself is transport‑agnostic, it can run over stdio, WebSockets, or HTTP.
- In practice, for Copilot Studio and Azure AI Foundry, you’ll usually expose it as an HTTP(S) endpoint so it’s accessible in the cloud.
- That’s why you see multiple hosting options:
- Local dev → run on your laptop, expose via a dev tunnel.
- Azure App Service / Container Apps → production‑ready, scalable.
- Azure Functions → serverless, event‑driven.
Refresh Date: 2025-11-03