Model Context Protocol (MCP).

Model Context Protocol (MCP) is a standard way for an AI application to communicate with external systems. An MCP server can expose defined capabilities, such as tools an application may call or information it may request. The AI application, acting as an MCP client, can use those capabilities through a consistent protocol rather than requiring a separate, custom connection pattern for every integration.

In Microsoft Copilot and Foundry contexts, MCP can be part of an integration design that connects an AI experience or application to business systems. MCP describes how the connection is structured. It does not, on its own, determine which Microsoft experiences support MCP, make an external system trustworthy, or decide what a user is permitted to access. Those details depend on the product and its configuration.

A typical interaction involves an AI application, an MCP client, and an MCP server that makes selected capabilities available. The client communicates with the server using the protocol; the server handles the connection to the underlying service. The AI application can then use the information or actions made available through that connection, subject to the controls in place.

The distinction between a tool and the authority to use it is important. MCP can provide a structured way to describe and invoke a capability, but authentication, authorization, approval, and data handling must be addressed by the surrounding systems. An organization should not assume that exposing a tool through MCP automatically makes its use safe or appropriately limited.

MCP is an integration concept, not a synonym for a particular Copilot feature or a guarantee of support across Microsoft products. Copilot experiences may differ in their intended use, Microsoft 365 application integration, and access to work data. For example, Copilot Chat and Microsoft 365 Copilot can have different work-data grounding and application integration, depending on the account, subscription, permissions, configuration, and service updates. MCP availability and behavior can also vary across experiences.

In Foundry-based application design, MCP may be relevant when teams want an AI application to interact with external tools or services through a defined interface. Whether a particular client, server, or deployment supports a given connection needs to be checked in that environment. The protocol can help shape an integration, but it does not remove the need to design for identity, data boundaries, reliability, and oversight.

Suppose an operations team wants an internal assistant to check the status of service requests. An MCP server could expose a narrowly scoped lookup tool connected to the organization’s request system. A user asks for the status of a specific request, and the assistant uses the available tool to retrieve the result.

The system still needs to determine who is asking, which requests that person may view, and whether the returned information is appropriate to show in the conversation. If the integration also allows updates, those actions should be treated differently from read-only lookups and may require additional safeguards or user confirmation.

MCP adds an integration boundary that should be governed like other connections between systems. A server may expose sensitive data or actions, and an incorrect configuration can broaden access beyond the intended use. The protocol itself does not replace an organization’s security model.

주요 고려 사항은 다음과 같습니다:

• Identity and permissions: Ensure access is tied to appropriate identities and limited to the data and actions each user or application needs.
• Tool scope: Expose only the capabilities required for the use case. Separate read-only access from actions that change business records.
• Trust and validation: Review the server, its implementation, and the systems it connects to. Treat returned data as input that may need validation.
• Monitoring and recovery: Plan how to detect failures or unexpected activity, and how to disable or roll back an integration if needed.
• Data handling: Consider what information may be sent to the server, returned to the AI experience, or retained by connected systems.

A measured rollout can help teams validate the connection before relying on it in business processes:

1. Define the task. Specify what users need the AI application to do and which systems contain the relevant information.
2. Choose the integration boundary. Decide which capabilities should be exposed through an MCP server, and keep the scope as narrow as practical.
3. Map identity and authorization. Determine how users or applications authenticate and how permissions are enforced across the connection.
4. Test the full interaction. Verify expected results, denied access, malformed requests, service failures, and any actions that modify data.
5. Deploy with oversight. Monitor usage and errors, document ownership, and establish a way to suspend the connection or revise its permissions.

The details of these steps depend on the chosen Microsoft experience, connected services, and deployment design.

MCP provides a structured protocol for connecting AI applications with external tools and data services. In Microsoft Copilot and Foundry scenarios, it can be useful as part of an integration approach, but its role and availability depend on the specific product and configuration.

The protocol does not replace identity controls, permissions, security review, or operational monitoring. Teams evaluating MCP should first define the task, limit exposed capabilities, test authorization and failure cases, and confirm that the intended Microsoft experience supports the required setup.