What is mcp server setup, and how does it work for AI agents?

### 1. What is the MCP Server Setup?

An MCP server starts with a mandatory handshake. The client, such as an AI agent or your Liminality router, sends an `initialize` request. This message includes the protocol version, client capabilities, and implementation details like name and version.

The server replies with its own version, capabilities, and info. It lists declared tools, resources, and prompts. The client then sends an `initialized` notification to confirm it is ready.

Configuration files handle the rest. `server.json` defines dependencies and execution arguments for distribution. Timeouts and server instructions usually depend on the SDK you use to build the server.

### 2. How Does It Work for AI Agents?

AI agents exchange data via JSON-RPC 2.0 over UTF-8 encoded transports. The flow is structured.

You can choose between stdio or Streamable HTTP. stdio uses standard input/output streams in a subprocess launched by the client. Streamable HTTP uses an MCP endpoint for POST requests. Responses come back as a single JSON object or a Server-Sent Events (SSE) stream. Clients can also start an SSE stream via HTTP GET.

The message patterns include standard Request/Response calls. Multi Round-Trip Requests (MRTR) allow the server to ask for more input. Subscribe/Notify handles event streams, letting the server push updates.

### Pending Action / Clarification Needed

These facts cover the generic MCP specification. You are building Liminality, a question-decomposition and routing MCP server.

To give you specific setup instructions for Liminality, I need two details. First, which transport mechanism are you using: stdio or Streamable HTTP? Second, which SDK are you using for the implementation? This choice dictates how you configure timeouts and server instructions.

Please confirm these so I can tailor the steps to your architecture.