MCP explained
MCP, the Model Context Protocol, is an open protocol through which an AI client reaches tools and data outside itself. The client is the application you talk to, for example Claude Code, Codex or Cursor. The server is whatever offers something to it: tools the model may call, resources it may read and prompts it may use.
How a client talks to a server
MCP knows two ways to connect. A local server runs as a program on your machine, and the client talks to it through standard input and output (stdio). A remote server is an address on the internet, and the client talks to it over HTTP, a transport called Streamable HTTP. A remote server usually wants to know who is calling: either through an OAuth sign-in in the browser, or through a token the client sends in a header.
OnlyMCP is a remote server. Its address is https://only-mcp.com/api/mcp, and it accepts both an OAuth sign-in and a token.
Why one address beats five
Without OnlyMCP every client needs every server configured separately. With five clients and four servers that is twenty entries, each with its own copy of a credential. With OnlyMCP it is five entries, one per client, all pointing at the same address.
- Connect a server once. It appears in every client that is connected to OnlyMCP, without touching a configuration file again.
- Credentials stay on the server. A client holds only its own OnlyMCP credential. The tokens for GitHub, Linear or your own server are stored encrypted by OnlyMCP and used on your behalf.
- Revoke one device, keep the rest. Every connected device is listed on its own and can be revoked on its own.
- The same names everywhere. A tool from a connected server keeps its prefix in every client, so
github__create_issuemeans the same thing on every machine.