What MCP actually is, if you have only seen the acronym
A plain explanation of the Model Context Protocol, what it replaced, and why it means one integration works in Claude, Cursor, OpenClaw and Hermes at once.
Optional — the post reads fine without one.
MCP stands for Model Context Protocol. Underneath the name it is unremarkable in the best way: JSON-RPC 2.0 over HTTP, with a handshake, a list of tools, and a way to call one.
That is genuinely all it is. The interesting part is not the protocol, it is what having any agreed protocol removes.
What it replaced
Before it, connecting a tool to an assistant meant writing an integration per assistant. The capability was the same every time — “post this” — and the plumbing was different every time.
With MCP you describe your tools once. Any client that speaks the protocol can discover them, read what each one does, and call it. We wrote our server once and it works in Claude, Cursor, Windsurf, OpenClaw and Hermes without a line of client-specific code.
What a server actually exposes
Three things, in our case: what accounts are connected, create a post, and read back how each network went. Each has a JSON schema describing its arguments, which is how the model knows that “schedule” takes an ISO timestamp or the word “now” without you explaining it.
The descriptions matter more than people expect. They are prompt, not documentation — the model reads them to decide what to call. A vague description produces a confidently wrong tool call.
Remote versus local
A local server is a process on your machine the client launches. A remote one is a URL the client calls, authenticated with a bearer token — which is what ours is, because a hosted product has nothing useful to run on your laptop.
The practical difference is setup: a remote server is a URL and a key in a config file, with no process to install, keep running or update.
The part worth being careful about
An MCP server that can publish under your name needs auth, and ours refuses any tool call without a valid API key — the handshake and tool list are open, because a client must be able to discover a server before it has been given a key.
The wider caution is about what you install. Skill and server registries have hosted malicious packages, and one audit put the rate at roughly one in twelve. Read the source of anything you connect to an account you care about. That advice applies to us as much as anyone: our skill file is a readable Markdown file in the repo, deliberately.