English for Model Context Protocol (MCP) Developers

The Model Context Protocol (MCP) has emerged as a standard way for AI applications to connect language models to external tools, data, and systems. If you work with MCP on an international team, you’ll need precise English to describe servers, capabilities, and the client-host architecture. This guide covers the core vocabulary for MCP developers.

Key Vocabulary

MCP server — a process that exposes tools, resources, or prompts to an AI application over a standard protocol, typically wrapping an external system or API. “We built an MCP server that exposes our internal ticketing system, so any MCP-compatible assistant can create and query tickets.”

MCP client — the component within a host application that connects to one or more MCP servers and manages the protocol communication. “The client discovers what tools each connected server offers and presents them to the language model as available actions.”

Host — the application that embeds an MCP client, orchestrating the conversation between the user, the model, and connected servers. “Our internal chat assistant acts as the host, managing connections to the ticketing, calendar, and knowledge-base MCP servers.”

Tool — a function exposed by an MCP server that the model can invoke to perform an action, such as creating a record or calling an API. “We exposed a create_ticket tool with clearly typed parameters, so the model knows exactly what fields are required.”

Resource — a piece of data an MCP server makes available for the model to read, such as a file, database record, or document. “We expose recent incident reports as MCP resources, so the assistant can read them for context without needing a dedicated tool call.”

Prompt (MCP primitive) — a reusable, parameterised prompt template that an MCP server can expose for clients to invoke with specific arguments. “The server exposes a summarize_incident prompt template that takes an incident ID and returns a structured prompt for the model.”

Transport — the underlying communication mechanism between an MCP client and server, such as stdio or HTTP with server-sent events. “We use the stdio transport for local development tools and HTTP for servers that need to run remotely.”

Capability negotiation — the initial handshake where a client and server agree on which protocol features (tools, resources, prompts) are supported. “During capability negotiation, our client learns that this particular server only supports tools, not resources, so it adjusts the UI accordingly.”

Sampling — a feature allowing an MCP server to request the host’s language model to generate a completion on its behalf, inverting the usual direction of control. “We use sampling so the server itself can ask the model to summarise a document, without needing its own separate LLM integration.”

Discussing Server Design

• “We kept our MCP server’s tool surface small and focused — five well-documented tools rather than twenty vague ones.”
• “Each tool’s description is written for the model to read, so we treat it like documentation, not just an internal comment.”
• “We separated read-only resources from state-changing tools, so the model’s read access doesn’t require the same approval as a write action.”

Talking About Architecture and Security

• “The host mediates every tool call, so we can add an approval step before any action that modifies production data.”
• “Capability negotiation means older clients degrade gracefully if a server adds a new feature they don’t yet support.”
• “We scope each server’s credentials narrowly, so even if the model is tricked into calling a tool unexpectedly, the blast radius is limited.”

Professional Tips

1. Write tool descriptions for the model’s audience, not just humans. Ambiguous descriptions lead to the model calling the wrong tool or supplying wrong parameters.
2. Explain the host’s mediating role when discussing safety. Stakeholders worried about “AI taking actions” are often reassured to learn every tool call passes through an approval layer.
3. Keep the tool surface minimal and composable. A server with fewer, well-defined tools is easier for both the model and human reviewers to reason about.

Practice Exercise

1. Explain to a colleague, in 3-4 sentences, the difference between a tool and a resource in MCP.
2. Write a short explanation (4-5 sentences) of why your team scoped an MCP server’s credentials narrowly.
3. Describe, in plain English, how capability negotiation lets an older client work with a newer server without breaking.