A program that exposes tools, resources, and prompt templates to Claude Code and other Model Context Protocol clients over a standardized protocol, so an agent can read from and write to external systems without bespoke per-integration code.
How it works
An MCP server advertises a fixed set of primitives to any compliant client: tools the agent can call, resources it can read, and prompt templates it can reuse. Client and server speak one standardized protocol to invoke them, so the integration to a system lives inside the server rather than inside the agent. Adding a new capability means standing up a server that advertises the new primitives; the agent learns about it at the protocol layer rather than absorbing the integration code.
Why it matters
Pinning the integration to a server makes a capability swappable and reviewable as its own unit, rather than tangled into the agent. The limit is the protocol vocabulary itself: the moment a capability needs a primitive the protocol does not define, MCP becomes overhead without payoff and bespoke integration is the honest answer. Portability holds inside the vocabulary, not beyond it.
In practice
A workflow that needs access to a database, a file store, and a search index connects to one server for each, and the agent calls their advertised tools through the same protocol rather than carrying three different integration paths.
Related standards and prior art
- Model Context Protocol: architecture · continuously updated defines the server primitives - tools, resources, prompts - over a JSON-RPC protocol
- WorkOS: MCP spec update · 2025-11-26 WorkOS write-up of the 2025-11-25 MCP spec update
Defined by Ready Solutions AI