Modern AI assistants are only as useful as what they can reach. MCP (the Model Context Protocol) is an open standard that lets an AI app plug into outside tools and data through one shared interface, instead of a bespoke integration for every service.
It is a simple client-server idea. An MCP client (Claude, Cursor, VS Code) connects to an MCP server that exposes three kinds of thing: tools (functions the model can call), resources (data it can read), and prompts (reusable templates). Click a capability below and watch a request travel across and back.
What just happened
Every message is JSON-RPC 2.0: a method (like tools/call), some params, and an id that ties each response back to its request. The server runs the call and returns a result. The model never needs to know how the server fetches data, only the shared shape of the conversation.
Two transports, same messages
The identical messages ride over one of two channels:
- stdio: the client launches the server as a local process and talks over standard input and output. Best for local tools on your own machine.
- Streamable HTTP: the server runs somewhere remote and speaks over HTTP (this replaced the older HTTP-plus-SSE transport). Best for hosted services many people can share.
Flip the transport toggle above and notice the messages don’t change, only where the server lives.
Why it matters
Before MCP, every AI app wired up every tool by hand. MCP makes tools portable: build one server, and any MCP-compatible client can use it. It is the same “standard socket” idea that made USB useful, a shared plug between models and the world’s data.
Connect your AI to this garden
The diagram above is a simulation, but this garden runs a real MCP server too. You can plug it into your own AI and ask it directly. It exposes two tools (search_notes, get_note) and every note as a note:// resource, over Streamable HTTP:
https://explorable-ai.vercel.app/api/mcp
Cursor (and other clients that speak HTTP), add to .cursor/mcp.json:
{
"mcpServers": {
"oliveaniss-garden": {
"url": "https://explorable-ai.vercel.app/api/mcp"
}
}
}Claude Desktop bridges a remote server through a local command:
{
"mcpServers": {
"oliveaniss-garden": {
"command": "npx",
"args": ["mcp-remote", "https://explorable-ai.vercel.app/api/mcp"]
}
}
}Then ask your assistant to search the garden for attention or read the note on embeddings, and it will call this server. That is a retrieval idea with a standard plug.