Glossary
Short answers to the terms you meet when you connect AI assistants to tools: one definition first, then the details, with official sources.
Updated 19 November 2026
- A2A (Agent2Agent)
- A2A (Agent2Agent) is an open protocol that lets AI agents built by different teams communicate and work together without exposing their internals to each other. MCP connects an agent to tools and data; A2A connects it to other agents.
- Agent skill
- An agent skill is a folder that teaches an AI agent how to do one kind of task. It holds a SKILL.md file (a name, a description of when to use it, and instructions) and can include scripts, templates and reference files. The agent loads it only when a task matches the description.
- AGENTS.md
- AGENTS.md is an open format: a plain Markdown file at the root of a repository that tells coding agents how to work in it, such as build and test commands and conventions. Many coding agents read it at the start of a session. CLAUDE.md is Claude Code's equivalent, and Claude Code can read AGENTS.md too.
- AI agent
- An AI agent is a system in which a language model decides, step by step, which tools to use and what to do next to finish a task, checking the result of each step before the next. A fixed script decides nothing; an agent directs its own process.
- Computer use
- Computer use is a capability that lets an AI model operate a computer the way a person does: it looks at screenshots and asks for mouse clicks and keystrokes, which your application carries out in an environment you control.
- Context window
- The context window is the amount of text a language model can look at while it writes a reply: your messages, files, tool results and its own earlier answers. It is the model's working memory, separate from what it learned in training, and it is measured in tokens.
- Enterprise-Managed Authorization
- Enterprise-Managed Authorization (EMA) is an MCP extension that lets an organisation grant access to MCP servers through its identity provider. Admins set the policy once, and people get the approved servers when they sign in, without a separate consent screen for each server.
- LLM
- An LLM (large language model) is an AI model with a very large number of parameters, trained on vast amounts of text, that generates text: it can answer questions, summarise, write and more. AI assistants are built on LLMs that are further trained to be helpful.
- MCP
- MCP (Model Context Protocol) is an open standard for connecting AI applications to outside tools and data. An AI app connects to small programs called MCP servers, asks each one what it offers, and calls those tools when a task needs them.
- MCP Apps
- MCP Apps is an official MCP extension that lets a tool return an interactive interface, such as a form, chart or picker, that the AI app shows inside the conversation instead of plain text.
- MCP host, client and server
- In MCP, the host is the AI application you use (such as Claude Desktop or VS Code), a client is the part of the host that keeps one connection to one server, and a server is the program that offers tools, resources and prompts. A host with three servers runs three clients.
- MCP Registry
- The MCP Registry is the official, central repository of metadata about publicly available MCP servers: each server's verified name, description, version and where to get it, stored in a server.json format. It points to code hosted elsewhere and is read mainly by directories and marketplaces.
- MCP server vs API
- An API is an interface that a program calls with code written for that API. An MCP server offers its capabilities through one shared protocol, describing each tool so that any MCP-compatible AI app can discover and call it without custom code. Many MCP servers are a thin layer over an existing API.
- Multi round-trip requests
- Multi round-trip requests (MRTR) are how an MCP server asks for more input in the middle of a call under the 2026-07-28 spec: it returns a result marked "input required", and the client calls again with the answers, instead of the server sending its own request over a held-open stream.
- OAuth for MCP
- OAuth is the standard way a remote MCP server gets permission to act on your account: you approve on the service's own sign-in page and the app gets a token for that server. Under the 2026-07-28 spec, apps identify themselves with Client ID Metadata Documents (CIMD) instead of Dynamic Client Registration.
- Prompt injection
- Prompt injection is an attack where instructions meant for an AI are hidden in content it reads, such as a web page, email, document or a tool's result, to make it do something the user didn't ask for, like leaking data or misusing a connected tool.
- RAG
- RAG (retrieval-augmented generation) is a technique where relevant documents are fetched from a knowledge base when you ask a question and passed to the language model with it, so the answer is grounded in that material instead of the model's memory.
- Remote vs local MCP server
- A local MCP server is a program the AI app starts on your own computer and talks to over standard input and output (stdio). A remote MCP server runs on someone else's machine and the app reaches it over the internet by URL, using Streamable HTTP and usually a sign-in.
- Streamable HTTP and stdio
- MCP has two standard transports. stdio connects an app to a server it starts on the same computer, through standard input and output. Streamable HTTP connects an app to a server over the network with HTTP POST, optionally streaming responses, and supports normal web authentication.
- Subagent
- A subagent is a specialised assistant that a main AI agent hands part of a task to. In Claude Code, each subagent runs in its own context window with its own instructions, tool access and permissions, so long side work stays out of the main conversation.
- Token
- A token is the smallest unit of text a language model works with: a word, part of a word, a character or a byte. Models read and write tokens, so context limits and API prices are counted in tokens rather than words.
- Tool annotations
- Tool annotations are optional hints an MCP server attaches to each tool: whether it is read-only, destructive, idempotent or reaches the open world. Apps use them to decide what to ask you before running a tool, but they are hints, not guarantees.
- Tool call
- A tool call is a structured request from an AI model to run one of the functions it was given, with arguments that match the function's input schema. The app runs the function, sends the result back to the model, and the model uses it to answer or to decide its next step.