MCP’s Three-Part Schema

▶ Watch (01:05)

MCP standardizes how applications give LLMs context. Think of it as a USB-C port for agents: one connection type, many endpoints. The protocol defines three primitives. Tools are callable functions, things like search or code generation. Resources are the data those functions act on, such as documents, calendar events, or support tickets. Prompts are templates that route user intent to the right tool. Together, these three form a single schema that any MCP-compatible agent can read and act on without custom glue code.

Copilot Studio Goes GA with MCP

▶ Watch (04:35)

Connecting an MCP server to Copilot Studio takes three steps: build the server with a supported SDK, create a custom connector from an OpenAPI schema file, then add it as a tool in your agent. Updates to the server reflect automatically. At Build 2025, MCP moved to GA.

“MCP in Copilot studio is now generally available.” — Zankar Desai

GA adds a visible tool list per server and enhanced tracing. He walked through the Copilot Studio lead-management MCP demo (06:49) live: “What are my open leads?” triggered one tool; a follow-up on a specific lead triggered a different one.

Zapier’s 30,000-Tool MCP Server

▶ Watch (09:13)

Zapier runs a hosted MCP server it manages entirely on its side. Authentication, authorization, and auditing all happen at Zapier. You copy your account ID, and every one of Zapier’s tools becomes available to your Copilot Studio agent.

“And we have over 8,000 apps that you can connect to,” — Nate Moore

That means 30,000 tools. Nate Moore showed the Zapier multi-app chain demo (11:02): one conversation found an urgent Gmail, created a Jira ticket, pulled LinkedIn profiles from a Google Sheet, updated the sheet with fresh data, and drafted a LinkedIn post. Each step used the previous turn’s context automatically.

Azure AI Foundry and Semantic Kernel

▶ Watch (17:04)

Azure AI Foundry Agent Service went GA this week. MCP support is not there yet, but Semantic Kernel bridges the gap today. Create a bare agent in Foundry, copy its ID, load the definition into Semantic Kernel, then attach an MCP plugin client-side. Mads Bolaris ran the Foundry + Semantic Kernel GitHub MCP demo (21:02), fetching the five latest Python issues from the Semantic Kernel repo. Foundry’s thread view exposed each tool call, the full JSON payload sent to the GitHub MCP server, and the response, useful for spotting and trimming oversized server output.

A2A vs MCP: When to Use Each

▶ Watch (26:57)

Google announced Agent2Agent protocol a few weeks before Build. A2A gives one agent a standard way to call another agent and hold a stateful conversation. MCP fits stateless tool calls: send a request, get a JSON payload back. A2A fits when the downstream system is itself an agent that carries context across turns. Mads walked through the A2A live call to Foundry agent (29:04): a C# client with no Microsoft packages beyond Azure Identity sent a question over A2A and got a live reply. SAP demonstrated a cross-vendor version routing Azure, Google, and SAP agents over the same protocol.

Notable Quotes

MCP in Copilot studio is now generally available. Zankar Desai · ▶ 5:37

And we have over 8,000 apps that you can connect to, Nate Moore · ▶ 9:38

Just FYI, Google did us dirty. Mads Bolaris · ▶ 30:57

Key Takeaways

  • MCP in Copilot Studio is GA, with tool-list visibility and enhanced tracing added at launch.
  • Zapier’s hosted MCP server gives Copilot Studio agents access to 30,000 tools across 8,000 apps with no setup.
  • Use MCP for stateless tool calls; use A2A when agents need to hold a multi-turn conversation with each other.