🥷🏻 How can I create my own MCP-like protocol?

Protocols are just the standards and rules defined by the internet community for communication at the network level.

MCP is a protocol that establishes secure communication between LLM applications and servers that expose custom executable function calls for your own business logic, as well as resources like files, databases, and media files.

The MCP protocol uses JSON-RPC 2.0 as its underlying protocol.

Every MCP call is an RPC (Remote Procedure Call) following a fixed request and response object format, similar to what we use in other applications.

Here is what a request format looks like:

{ "jsonrpc": "2.0", "id": 1, "method": "add", "params": [1, 2] }

And the server responds to the request in the following format:

{ "jsonrpc": "2.0", "id": 1, "result": 3 }

Now let’s understand what MCP adds on top of this.

MCP hosts incorporate an MCP client with an LLM instance, establishing a 1:1 connection with the MCP server.

MCP hosts with LLM calls are able to understand from the prompts what tool calls are required to fulfill the prompt’s output, along with the general LLM results.

The list of tools captured by the MCP client provides the available tool objects, including details like tool name, description, argument schema, and their RPC handler.

Once the MCP host identifies the list of qualified tools (based on their descriptions, names, and arguments) that can be called, the execution handlers receive the data from the LLM—such as tool names and arguments—which can then be parsed into RPC calls as shown in the request object above.

Similarly, you can define handlers to access your static resources and provide their paths through RPC calls.