Building a No-Code MCP Server

A no-code or low-code workflow platform can expose workflow actions as MCP tools. This approach makes it possible to create a basic server without writing the protocol implementation directly.

A simple server can begin with an MCP server trigger. Tools are then attached to the trigger. One example adds a calculator and an email-sending tool.

The calculator tool provides the ability to evaluate a mathematical expression. The email tool connects to a Gmail account through configured credentials. Its operation is set to send a message, and fields such as the recipient, subject, and message can be left for the model to decide from the user's instruction. The message format can be set to HTML.

After the workflow is saved, the server is activated. The platform provides a production URL that allows a client to connect to the server.

A second workflow can act as the host and client. It contains an AI agent, a chat model, and an MCP client tool. The production URL from the server is entered as the client endpoint. Streamable HTTP is selected as the transport.

A calculator request is a simple way to verify the connection. When the user asks the agent to calculate 11 multiplied by 99, the agent passes the request to the MCP client, the client invokes the calculator tool on the server, and the result returns to the model. The final answer is 1,089.

The same server can then be used for an email request. The user gives the recipient and purpose in natural language. The model fills the relevant email fields and invokes the server tool. In a desktop host, the user sees that the application wants to use the email function and can allow the call once before the message is sent.

The server can be moved between hosts. To connect it to a desktop AI application, the server definition and endpoint are added to the application's configuration, the application is restarted, and the tools appear in the host's tool list. The host can then use the same calculator and email functions.

The no-code approach is accessible and fast, but the described implementation has an important limitation: it can expose tools but does not provide the same ability to define additional resources and prompt templates. A coded server is needed when the implementation must expose all three MCP feature types.