Claude Code Glossary

The configuration file that tells Claude Code which MCP servers to launch, with what command, arguments, and environment variables. It can live at the project level so a team shares the same servers.

Anthropic's command-line coding agent. It runs in your terminal, understands your project, and can read files, run commands, and edit code to accomplish a task you describe in natural language.

A markdown file in your project (or home directory) that Claude Code loads into every session as standing instructions - your conventions, commands, and rules. Because it loads every time, keeping it tight saves tokens.

The act of summarizing a long conversation to free up the context window while preserving the important state. Compacting lets a long session continue without reloading its entire history each turn.

The total amount of text (measured in tokens) the model can consider at once - your prompt, the files it has read, the conversation history, and its own output. When it fills up, older content must be dropped or compacted.

Running Claude Code non-interactively, driven by flags and piped input rather than a live conversation. It is how you slot the agent into scripts, CI pipelines, and automation.

A command Claude Code runs automatically at a defined moment in its lifecycle - for example before or after a tool call. Hooks let you enforce rules, run linters, or block actions deterministically, outside the model's judgment.

Model Context Protocol - an open standard for connecting AI tools to external systems. An MCP server exposes tools, resources, or data (a database, an API, a service) that Claude Code can then use.

A program that implements the Model Context Protocol to give Claude Code new capabilities - reading from a database, calling an API, or accessing a service. You configure which servers to load in .mcp.json.

The three tiers of the Claude 4 family. Opus is the most capable and costly, Sonnet is the balanced default, and Haiku is the fastest and cheapest. You can switch between them per task.

A configurable profile that shapes how Claude Code communicates - its tone, verbosity, and formatting - so responses consistently match how you like to work.

Rules that control which tools and commands Claude Code may run. 'Allow' runs without asking, 'ask' prompts you for approval, and 'deny' blocks outright - the core of running the agent safely.

A mode where Claude Code proposes a plan and waits for your approval before making changes. It is a safety and alignment step - you catch a wrong direction cheaply before any code is written.

Two of the hook events. A PreToolUse hook runs before a tool executes and can allow, block, or modify it; a PostToolUse hook runs after and can react to the result. They are how you gate or audit tool activity.

A mechanism that stores repeated context (like your CLAUDE.md or unchanged file contents) so it can be reused across prompts at a large discount instead of being re-processed at full price each turn.

The Claude Code configuration file (in .claude/) that holds permissions, environment variables, model choice, hooks, and other options. Project and user settings combine to control how Claude Code behaves.

A packaged capability, defined in a SKILL.md file with frontmatter, that Claude Code loads and can invoke when its trigger conditions match. Skills bundle instructions (and optionally scripts) for a repeatable task.

The markdown file that defines a skill. Its frontmatter sets the name and description (which control when the skill triggers), and its body holds the instructions the skill runs.

A reusable, named command you invoke with a leading slash (like /review). Custom slash commands are defined as markdown files in .claude/commands and expand into a prompt, optionally taking arguments.

A customizable line in the Claude Code interface that can display context like the current model, git branch, directory, or token usage, driven by a command you configure in settings.

A specialized, separate agent Claude Code can delegate a scoped task to. Each subagent has its own instructions and its own context window, so it can do focused work without cluttering the main conversation.

The high-level instructions that set the model's role and behavior for a session. In Claude Code you can extend it (for example with --append-system-prompt) to add standing guardrails on top of the defaults.

The unit models read and write in. A token is roughly a few characters - often about three-quarters of a word in English. You are billed per input token and per output token, which is why token count drives cost.

When the model invokes a capability - reading a file, running a command, searching, or calling an MCP tool - rather than just producing text. Agentic work is largely a sequence of tool calls.

How an MCP server communicates with Claude Code. Stdio runs the server as a local subprocess over standard input/output; SSE and HTTP connect to a server over the network. A mismatch is a common cause of connection failures.