1. Why another protocol?

If GPT-4 was the brain, and Azure AI Foundry is the nervous system, MCP is the USB-C port: a tiny, universal socket that lets your agent grab new super-powers from anywhere on the internet. That was Microsoft’s sound-bite at Build 2025, where it joined the MCP steering committee and pledged first-party support across GitHub, Copilot, Dynamics 365, Windows 11 and now Azure AI Foundry.

Until now, Foundry agents could only call the built-in tools Microsoft shipped—or whatever you wrapped in a custom function. MCP smashes that wall: any service that speaks the open-source Model Context Protocol can be wired in with two parameters and zero code changes.

2. What exactly is MCP?

• Open standard (Apache 2.0) that describes how LLM hosts discover, describe, and safely invoke external tools.
• JSON-RPC 2.0 over your choice of transport (STDIO, SSE, WebSockets, HTTP/2).
• Client-server topology:
  • MCP host = your agent
  • MCP server = a lightweight process exposing one or more “tools” (functions with a JSON schema)
  • MCP client = a connector that brokers messages between them.

Think “function calling on steroids”—multi-call chains, streaming arguments, and granular permission scopes baked in.

3. New in Azure AI Foundry (27 Jun 2025)

Capability	Details
MCP tool type	Add {"type":"mcp"} to an agent; one entry per remote server.
Bring your own server	Point to any MCP endpoint—GitHub, Morningstar, Neo4j, internal micro-service, you name it.
Custom headers	Pass API keys or OAuth tokens at run time via tool_resources.headers.
No approval prompts	Preview supports require_approval: "never" only, so review endpoints carefully.
Supported regions	westus, westus2, uaenorth, southindia, switzerlandnorth.
SDK parity	Works today in REST, Python, .NET, and JavaScript SDKs.

4. Five-minute Quick-Start

1. Spin up (or reuse) a Foundry agent bashCopyEditaz foundry agent create \ --name my-agent \ --resource-group rg-demo \ --project standard
2. Register your first MCP server (e.g., GitHub) jsoncCopyEdit# PATCH body for agent update { "tools": [ { "type": "mcp", "server_url": "https://api.githubcopilot.com/mcp/", "server_label": "github", "require_approval": "never" } ] }
3. Invoke a run and pass auth headers jsoncCopyEditPOST /agents/my-agent/runs { "tool_resources": { "tool_label": "github", "headers": { "Authorization": "Bearer <gh-token>" } }, "messages": [ {"role":"user","content":"Create a PR on repo contoso/api that bumps AI SDK to 3.0"} ] }
4. Watch the magic – the agent calls the GitHub MCP server, triggers the Pull Request, then replies with a link—all in one chat turn.

(Full payload samples live in the docs sandbox.) learn.microsoft.com

5. Security & governance checklist

• Treat the server as untrusted code → only use endpoints you own or vendors you’ve vetted.
• Log everything → headers are ephemeral but payloads could still contain sensitive data.
• Network boundaries → lock servers behind VNet + Private Link where possible.
• Least-privilege tokens → scope API keys to the smallest set of actions the tool needs.

Microsoft’s preview docs hammer these points precisely because all traffic ultimately leaves Azure’s boundary.

6. Early use-cases we’re seeing

Scenario	Example MCP server
Vector search RAG	https://embeddings.acme.ai/mcp exposes similarity_search against your private PGVector index.
Finance analytics	Morningstar-MCP returns getFinancials(ticker,period) for real-time ratios.
DevOps copilots	GitHub or Azure DevOps MCP servers manage PRs, pipelines, issue triage.
Knowledge graphs	A Neo4j MCP server surfaces Cypher queries as tools—agents reason over graph data effortlessly.

The beauty: you can chain multiple servers in one agent, fusing data from GitHub and Morningstar in a single answer without writing orchestration code.

7. Current limitations & roadmap clues

• Approval workflows – no “first-call consent” yet, expect UI / policy hooks later in preview.
• Latency – first call includes schema discovery; cache results for P99 improvements.
• GUI tooling – Visual Studio Code extension shows MCP servers in the tree view but lacks click-to-add wizard (on the backlog according to internal demo at Build).

Microsoft’s pattern is clear: ship raw API first, then layer polished UX, so jump in early if you want feedback influence.

8. Where to dig deeper

• Docs: Connect to MCP Servers (Preview) in Learn.
• Spec: <modelcontextprotocol.io> (open governance on GitHub).
• Sample code: Tech Community blog post with a Chainlit + Azure OpenAI end-to-end repo.
• Community: #mcp channel in the Azure AI Discord.

Final thoughts

The MCP tool turns Azure AI Foundry into a genuine agent platform, not just a chat wrapper. By externalising “skills” as network-addressable modules, teams can ship tiny, domain-specific services and let agents compose them on demand—much like micro-services revolutionised backend design a decade ago.

If you’ve been mapping out complicated function-calling glue code, rip it out, give MCP a spin, and tell us what you build. The port is finally open. 🚀