MCP (Model Context Protocol) is an open protocol that defines how an AI agent connects to tools and data. Your organization sets up an MCP server that exposes a closed list of actions and data to the agent, with a clear description of each one. The agent sees only what was defined and works within its permissions. One server serves several agents, so you build the connection to SAP, Priority ERP or Dynamics 365 once and reuse it.
What is MCP, in plain terms?
MCP stands for Model Context Protocol. It is an open protocol that Anthropic introduced in November 2024, and since then many platforms and tools have adopted it, including Microsoft's Copilot Studio and Azure AI Foundry. The protocol sets a common language between an AI agent and the systems it needs to work with.
It has two sides. The client (MCP Client) sits inside the application that runs the agent. The server (MCP Server) sits next to the business system and mediates access to it. The server exposes three kinds of components:
- Tools: actions the agent can run, such as "find a supplier by number" or "create an invoice draft". Each tool has a name, a description and defined parameters.
- Resources: data the agent can read, such as a record, a document or a table.
- Prompts: ready-made templates the server offers for reuse.
When the agent connects, it receives the list of tools and their descriptions from the server. During the conversation, the model decides which tool fits the request, and the server runs the action against the system and returns a response. The server can run on the same machine or over the network, through HTTP.
Why does MCP matter for enterprises?
Without a common protocol, every new agent needs its own connection to every system. After a few agents, the organization has many similar connections, each with different permissions and logic. MCP brings order to this:
- One integration layer: you build one MCP server per system, and several agents use it. A new agent connects to the existing server.
- Defined actions instead of open access: the agent does not get access to the database. It gets a closed list of actions that the organization chose.
- Permissions in one place: what may be read, what may be written, and by whom. You change it in the server, and it applies to all agents.
- Audit and control: every tool call goes through the server, so you can log it and check who asked for what and when.
- Less dependence on a single platform: the protocol is open, and the same server can serve an agent in Copilot Studio, an agent in Azure AI Foundry or another tool that supports MCP.
How is MCP different from an API call, a connector or a plugin?
MCP does not replace an API. Usually the MCP server itself calls the system's API. The difference is in who decides which action to run, and who can use the connection.
| Direct API call | Power Automate connector or plugin | MCP server | |
|---|---|---|---|
| What it is | Code that calls a system endpoint | A ready-made component that wraps an API inside a specific platform | A server that describes to the agent which tools and data are available, and runs them |
| Who decides what to run | The developer, in advance, in code | Whoever built the flow or the agent on the platform | The agent, from a closed list the server exposes |
| Who can use it | The application the code was written for | Flows and agents on the same platform | Any agent or application that supports MCP |
| Reuse | Low: each project writes the connection again | Within the platform | High: one server serves several agents |
| When it fits | A fixed process that is known in advance | Automation in Power Platform against systems that have a connector | An agent that needs to choose an action on its own, against one system or several |
In practice, you combine them. Whatever is fixed and known in advance, such as a rule-based calculation, stays in a defined flow. The agent's access to the system goes through defined tools in an MCP server. At Gil Car, for example, an MCP layer connects the agent to SAP, and four dates are calculated in Power Automate.
How do you keep it secure when an agent accesses core systems?
MCP is a protocol, not a security mechanism. Security depends on how you build the server. These are the principles we work by:
- Least privilege: the server exposes only the entities and actions the process needs. Not "access to SAP", but "read supplier" and "create draft".
- Read separate from write: read tools and write tools are defined separately, with separate permissions.
- Writes as drafts or with approval: a write action creates a draft, or waits for a user's approval. The final posting stays with an authorized person.
- Validation before every action: the server checks the data against the system, for example to prevent duplicates, before it does anything.
- User identity and permissions: the agent acts on behalf of an identified user and within that user's permissions. For servers that run over the network, the MCP specification defines OAuth-based authorization.
- Full logging: every tool call is logged, with its parameters and result.
- On-premises systems: when the ERP sits behind a firewall, the server runs inside the organization's network or reaches it through a gateway, without exposing the system to the internet.
Where does MCP run in the Microsoft world?
Copilot Studio supports connecting to MCP servers as agent tools. You add the server, and the agent gets the tools it exposes. When you add a new tool on the server, it becomes available to the agent without rebuilding the connection. This fits internal agents in Teams and Microsoft 365 that need to work with the ERP. Read more on Copilot Studio agent development.
Azure AI Foundry lets you connect an agent to an MCP server as a tool, alongside custom code and other tools. This fits more complex agents, or agents that serve customers.
The same MCP server can serve both platforms. We covered the differences between them in the guide Copilot Studio vs. Azure AI Foundry.
Examples from Zooz-Tech projects
Gil Car · Invoice agent on SAP on-premises: an agent in Copilot Studio reads supplier invoices, checks them against existing records in SAP to prevent duplicates, and creates a draft in JournalVouchers through a dedicated MCP layer. Four dates are calculated in Power Automate. The final posting happens only after an authorized user has reviewed and approved it inside SAP.
Moroccanoil · Export on Dynamics 365 F&O: automation that turns a single Packing List email into a complete export document package in Dynamics 365 F&O, and an AI agent that answers free-form questions about the data through an MCP layer, without direct access to the database.
Read more about these integrations on SAP & Priority integration.
How do we build an MCP layer?
This is how it works in our enterprise AI agent projects:
- Start from the process: what the agent needs to read, what it needs to do, and where a person's approval is required.
- Define the tools: a name, a clear description and parameters for each action. A good description helps the model choose the right tool.
- Build the server against the system's API, with validation, permissions and logging.
- Connect the agent in Copilot Studio or Azure AI Foundry, and test it on real scenarios, including requests it must not carry out.
- Go to production gradually: reading first, then writing as drafts with approval.
- Add tools and processes to the same server, and let more agents use it.
Frequently asked questions
Does MCP replace the system's API?
No. The MCP server usually calls the system's API. It adds a layer on top that describes to the agent which actions are available, and limits it to those actions only.
Is MCP a Microsoft technology?
No. MCP is an open protocol that Anthropic introduced in November 2024. Microsoft supports it in Copilot Studio and Azure AI Foundry, and many other platforms support it as well.
Can you connect an AI agent to SAP on-premises through MCP?
Yes. At Gil Car, Zooz-Tech built an MCP layer that lets an agent in Copilot Studio check against SAP on-premises data and create a draft in JournalVouchers. An authorized user approves every draft inside SAP.
Can the agent change data in the system without approval?
Only if you give it a tool that does that. We define write actions as drafts or with user approval, and every tool call is logged.
Can several agents use the same MCP server?
Yes. One server can serve several agents, including on different platforms. Permissions are defined in the server, so they apply to all of them.
CTO at Zooz-Tech. Leads development and architecture of AI agents, Generative UI and Azure infrastructure.