MCP is a USB port for AI tools. Hosted MCP means someone else runs the adapter.
When X says “the hosted X MCP is free to connect to,” it means: X runs a Model Context Protocol server so tools like Grok, Claude, or Cursor can call the X API without you deploying anything. You bring OAuth credentials; API calls are pay-per-use. Here's the whole machine, in 3D.
Drive the protocol
The three roles
Host
The AI app you're using — Grok, Claude Desktop, Cursor. It decides when a tool is worth calling based on your conversation.
MCP client
A small connector inside the host. Speaks JSON-RPC 2.0, maintains one session per server, relays tool schemas to the model.
MCP server
Wraps a capability — the X API, a database, your filesystem — and exposes it as typed tools, resources, and prompts.
The point of the standard: any host can use any server. Before MCP, every AI app wrote bespoke plugins for every API — an N×M integration mess. With MCP it's N+M: X writes one server, every MCP-speaking tool can use it.
Hosted vs. self-hosted MCP
| Hosted (e.g., X's MCP) | Self-hosted | |
|---|---|---|
| Who runs the server | The API provider, on their infra | You, locally or on your VPS |
| Setup | Point your client at the URL + OAuth app | Install, configure keys, keep it updated |
| Cost to connect | Free; underlying API calls pay-per-use | Your compute + API costs |
| Trust model | Provider sees your requests | Traffic can stay on your machine |
| Best for | Public APIs (X, Stripe, GitHub) | Private data: files, internal DBs |
For the X case specifically: you create a free X Developer app to get OAuth credentials, connect your client to the hosted endpoint, and pay only for the API calls you make (credit-based, no subscription). The MCP layer itself adds no fee — it's distribution, not a product.