// glossary
Model Context Protocol plain-English.
Operator-grade definition. Plain words, plus where the term shows up in real work.
What is a Model Context Protocol?
Model Context Protocol — MCP — is an open standard for connecting LLM hosts to external tools, resources, and prompts. It defines the wire between an AI host and a server that exposes capabilities: discovery, schema, invocation, streaming, cancellation, and notifications.
MCP exists because the integration problem between models and the systems they need to act on was repeated in incompatible ways at every host. Each model app — Claude, Cursor, Codex, Copilot — had its own plugin or tool-calling shape. MCP is the standardization layer so one server can serve many hosts.
The protocol is intentionally narrow. Servers expose three things: tools (callable functions with typed schemas), resources (read-only context the model can reference), and prompts (parameterized templates). The host handles model orchestration, user consent, and policy; the server handles the actual work.
Why it matters in production: a standardized protocol means observability, auth, rate limiting, and audit can be reasoned about generically — instead of reinvented per host. That is what turns AI from one-off integrations into infrastructure.
Related questions
Is MCP open source?
Yes — open specification with reference SDKs in TypeScript and Python, plus a growing community of servers. The protocol itself is free to implement.
Which hosts support MCP?
Claude (desktop and Code), Cursor, Codex, and a growing set of agent runtimes. Host support is the reason it earned standardization status quickly.
MCP vs OpenAPI: are they competing?
Different layers. OpenAPI describes HTTP APIs. MCP describes the host-to-tool wire with model-specific affordances (prompts, resources, streaming). They compose — an MCP server often fronts OpenAPI-described services.
Related work
Need someone to actually run this in production? book a call.