MCP Protocol Specification (modelcontextprotocol.io)

The Model Context Protocol (MCP) is a new open specification that standardizes how LLM applications connect to external data sources and tools. Built around a TypeScript schema (schema.ts) and using JSON-RPC 2.0 for stateful, capability-negotiated connections, MCP defines three roles—Hosts (LLM apps), Clients (connectors inside hosts), and Servers (context/capability providers)—and a set of features servers can expose (Resources, Prompts, Tools) plus a client-offered Sampling feature for server-initiated agentic behaviors. The spec also prescribes utilities such as configuration, progress reporting, cancellation, error logging, and treats RFC 2119 keywords (MUST/SHOULD/etc.) as normative where capitalized. Implementation examples live at modelcontextprotocol.io. MCP’s significance lies in bringing LSP-style standardization to AI integrations: it promises composable workflows, reusable connectors for IDEs, chat UIs, and custom pipelines, and clearer interoperability for tool invocation and recursive model interactions. At the same time MCP explicitly grounds security and trust-safety: user consent and granular control, strict data-privacy rules, cautious handling of arbitrary code execution, and explicit LLM-sampling approval are required. Because MCP enables broad data access and tool execution, implementors must build robust consent/authorization flows, access controls, prompt-visibility limits, and document security trade-offs before deploying connectors in production.