MCP servers explained for developers (2026).

The shape of MCP

MCP has three concepts:

• Tools. Functions the assistant can call. Each has a name, a JSON-Schema input, and a structured output. This is the most-used MCP feature in 2026.
• Resources. Read-only data the assistant can request by URI: files, database rows, config. The client pulls them; the server serves them.
• Prompts. Server-defined prompt templates the assistant can use directly. Less common in practice; mostly used by domain-specific helpers.

For a deeper breakdown of which primitive to reach for, see resources vs tools vs prompts.

Transports

Two transports, picked by the client:

• stdio: the client spawns the server as a child process and communicates over stdin/stdout. Used by CLI assistants (Claude Code, Codex, Cursor). Auth comes from the parent process.
• Streamable HTTP: the client POSTs to the server over HTTPS and reads responses as Server-Sent Events. Used by hosted assistants (Claude.ai, ChatGPT custom connectors). Auth is OAuth 2.1 with Dynamic Client Registration (RFC 7591).

Why every assistant adopted it

Pre-MCP, each AI assistant had its own tool-call shape: OpenAI's function calling, Anthropic's tool_use, Google's function-declarations, etc. Tool authors had to ship N integrations to reach N assistants, and rewrite each when the host changed its calling convention. MCP collapses that into one tool surface that every assistant agrees to read.

The adoption curve in 2025–2026 was steep because the alternative was unsustainable. Anthropic open-sourced the spec, OpenAI integrated it into Codex CLI, Cursor adopted it natively, Microsoft added it to GitHub Copilot, and the community shipped ~500 servers in the first year.

A minimal MCP server in 50 lines

// server.ts: TypeScript SDK
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

const server = new McpServer({
  name: "my-server",
  version: "1.0.0",
});

server.tool(
  "say_hello",
  "Greet someone by name",
  { name: z.string() },
  async ({ name }) => ({
    content: [{ type: "text", text: "Hello, " + name + "!" }],
  }),
);

await server.connect(new StdioServerTransport());

Ship it on npm; users add it to their assistant of choice via one line of config.

Where MCP is in 2026

• Server registry. modelcontextprotocol/servers on GitHub catalogues the official + community list.
• Spec maturity. The protocol has settled and breaking changes are rare, though the 2026-07 release candidate is the largest revision yet: a stateless core, response caching, and a formal deprecation policy. Recent additions include authorization (OAuth 2.1) and elicitation, where a server can prompt the user mid-call.
• Discovery. Most clients now have a UI for browsing and installing MCP servers, including OAuth-protected hosted ones.
• Production usage. Real teams ship internal MCP servers for their custom systems, which is where the protocol stops being a curiosity and starts being infrastructure.

Where Stacktree fits

stacktree-mcp is the MCP server behind Stacktree. The easiest way to wire it into any coding agent is npx stacktree-install. Pick Claude Code, Cursor, Codex, OpenCode or Amp during the prompts and you get seven tool calls in that agent: publish HTML, replace HTML in place, gate by password or email domain, set expiry, list, delete. The protocol is the whole story: there's no separate API to learn, no separate auth model, no separate transport. If your assistant speaks MCP, it speaks Stacktree.