What is MCP?

MCP is a specification for connecting language model applications to the outside world. Instead of every application writing its own integration for every data source or tool, MCP defines one message format, based on JSON-RPC 2.0, that both sides speak. The specification says it took inspiration from the Language Server Protocol, which did a similar job for programming-language tooling.

Three roles take part. A host is the LLM application that starts connections. A client is a connector that lives inside the host. A server is a service that provides context and capabilities. Servers can offer resources (data for the user or the model), prompts (templated messages for users) and tools (functions the model can run). Clients can offer features back to servers, such as elicitation, which lets a server ask the user for more information.

This entry follows the 2026-07-28 revision of the specification. In that revision requests are stateless and self-contained, and each one carries its own capability declaration. Optional extensions, such as Tasks for long-running work, are negotiated separately and only apply when both sides support them.

The specification is candid about one limit: the protocol cannot enforce its own security principles. Hosts and implementors are expected to build consent and authorization flows around tool calls and data access.