Introduction
When you build an agent powered by the GitHub Copilot harness in Microsoft Copilot Studio, some requests need real-time data or an action in your organization's own systems. These systems can include custom services, databases, and internal APIs that no prebuilt connector covers. A Model Context Protocol (MCP) server can expose capabilities from those systems to your agent as tools. You need to connect the agent to the server, choose which of the server's tools the agent uses, and confirm that the agent calls the right tool.
Suppose a platform team in your organization runs one central MCP server. The server exposes tools for several internal systems, such as customer records and loan application status. The platform team owns the server and provides its URL and the credentials you need to connect. You're building a customer-service agent that needs some of those tools, but not all of them. When a customer asks about a loan application, the agent needs to call the right tool and return a current answer.
In this module, you explore:
- When to use an MCP server instead of a connector for a custom or internal service
- How to add an MCP server to an agent and configure the authentication the server requires
- How to choose which MCP tools the agent uses
- How to test MCP tool calls and troubleshoot calls that fail or don't run
By the end of this module, you'll be able to connect an agent powered by the GitHub Copilot harness to an MCP server, choose which of the server's tools the agent uses, and verify that the right tools run.
Note
Adding an MCP server to an agent powered by the GitHub Copilot harness is currently in preview. Features and the user interface can change.