One Protocol To Plug Them All
If you've been living under a rock, the Model Context Protocol just won the connector war and nobody held a vote. Anthropic shipped it as an open spec, and within twelve months Cursor, Zed, Claude Desktop, Cline, Windsurf, and roughly every other tool with an LLM in it adopted MCP. Spoiler: this is the USB-C moment for AI.
The Setup
Before MCP, every editor invented its own custom tool-call format. ChatGPT had plugins, Cursor had its own thing, Claude Desktop had another. Building one integration meant writing it five times. MCP collapses that into a single JSON-RPC contract over stdio or HTTP.
{
"name": "rebuild-relief-pipedrive",
"version": "0.1.0",
"bin": { "rr-pipedrive": "./dist/server.js" },
"dependencies": {
"@modelcontextprotocol/sdk": "^1.0.0"
}
}The Money Pattern
An MCP server exposes resources, tools, and prompts. The client (Claude, Cursor, whatever) discovers them at connect time and exposes them to the model. The model calls a tool, the server runs it, the result streams back. That's the whole protocol.
server.tool("get_deal", { id: z.string() }, async ({ id }) => {
const deal = await pipedrive.deals.get(id);
return { content: [{ type: "text", text: JSON.stringify(deal) }] };
});The Catch
Tooling is still rough. The MCP Inspector exists but debugging stdio transports feels like 2008 — you're tailing logs, eyeballing JSON, and hoping the schema validates. Auth is also a half-finished story. The HTTP transport with OAuth is shipping but the ecosystem hasn't caught up yet.
The Verdict
MCP is the closest thing AI has had to a real standard since OpenAI's chat completions API became the de facto interface. I'm rewriting the Rebuild Relief Pipedrive + BigQuery integrations as MCP servers right now because once you have the connector once, every editor and agent on the planet can use it. Bet on the protocol. Do not @ me when LangChain pivots to it next quarter.