Two protocols. One playground.

REST (REpresentational State Transfer) is an architectural style for building networked applications, introduced by Roy Fielding in his 2000 dissertation. It is not a protocol or a standard — it's a set of constraints that, when followed, produce systems that are simple, scalable, and reliable.

Most "REST APIs" you'll meet in the wild are HTTP APIs that follow most of these constraints. The pattern is the same everywhere: a client sends an HTTP request to a URL identifying a resource, using an HTTP method that describes the intent, and gets back a representation (usually JSON).

GET /api/users/42 HTTP/1.1
Host: api.example.com
Accept: application/json
Authorization: Bearer eyJhbGciOi...

HTTP/1.1 200 OK
Content-Type: application/json
Cache-Control: max-age=60

{
  "id": 42,
  "name": "Ada Lovelace",
  "email": "ada@example.com",
  "_links": {
    "self":   { "href": "/api/users/42" },
    "orders": { "href": "/api/users/42/orders" }
  }
}

Two concepts trip people up constantly. They're easy once you have a mental model.

MCP is an open standard, introduced by Anthropic on November 25, 2024, for connecting AI applications to external systems — data sources, tools, and workflows. Think of it as USB-C for AI: one consistent plug, hundreds of compatible devices.

Without MCP, every AI application needs custom integrations for every external system it wants to use. With MCP, an AI app and a service speak the same protocol, so you "build once and integrate everywhere". The protocol takes direct inspiration from the Language Server Protocol (LSP), which solved the same M-by-N integration explosion for code editors and language tooling.

Adoption since launch: OpenAI (March 2025), Google DeepMind (April 2025), plus Microsoft, GitHub, Cursor, Sentry, Zed, Replit, and hundreds of community servers. The current stable spec is 2025-11-25; a release candidate 2026-07-28 introduces a stateless transport and server-rendered MCP Apps.

MCP uses a client-server architecture with three participants. A single AI application (the host) can speak to many MCP servers at once, opening one dedicated client per server:

MCP has two conceptual layers:

• Data layer — the JSON-RPC 2.0 protocol: lifecycle, capability negotiation, primitives (tools, resources, prompts), and notifications.
• Transport layer — how messages travel: stdio for local subprocesses, Streamable HTTP for remote servers (HTTP POST plus optional Server-Sent Events).

Client primitives

Servers can also ask the client to do things:

The same JSON-RPC messages flow over one of two transports. Local or remote — that's the choice.

// Host launches server, then talks via stdin/stdout
{
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/docs"]
}

POST /mcp HTTP/1.1
Host: mcp.sentry.io
Authorization: Bearer sk_live_...
Content-Type: application/json
Accept: application/json, text/event-stream

{"jsonrpc":"2.0","id":1,"method":"tools/list"}