Glossary

Claude Code has accumulated a fair bit of vocabulary. This glossary gives each term a short, practical definition and points to the handbook page that covers it properly. Model-level ideas such as tokens or temperature are not covered here; this is about the tool.

A

AGENTS.md

A Markdown instructions file aimed at AI coding agents in general rather than Claude specifically. If a repository has AGENTS.md but no CLAUDE.md, Claude reads it as the project instructions, so teams using several tools need only one file. The Project instructions option in /config lets you read both files or only CLAUDE.md. Direct AGENTS.md support needs v2.1.277 or later; in sessions where it is unavailable, import it from a CLAUDE.md instead. See Memory.

Agent teams

Several separate Claude Code sessions working together under a lead session, sharing a task list and able to message one another. Unlike subagents, each teammate has its own context window and you can talk to any of them directly. Experimental and off by default. See Agent teams.

Agentic coding

Working with an AI that can act (read files, run commands, edit code) rather than one that only suggests text for you to paste. Claude Code qualifies because it has tools. See How Claude Code works.

Agentic harness

Everything wrapped around the model that makes it useful as a coding agent: file access, shell execution, permission checks, memory loading and the loop that chains steps together. Claude is the model; Claude Code is the harness.

Agentic loop

The repeating cycle behind every task: gather context, act, check the result, go again until done. Each tool result feeds the next decision, and you can interrupt at any point. Hooks, skills and MCP all attach to particular stages of it.

Artifact

A private, interactive web page that a session publishes to claude.ai, useful when output reads better visually than in a terminal or needs sharing. Republishing updates it in place, and it lives in the same gallery as artifacts from claude.ai chats. See Artifacts.

Auto memory

Notes Claude keeps for itself about your preferences and corrections, stored per git repository under ~/.claude/projects/ and shared across that repository's worktrees. The first 200 lines or 25 KB of its MEMORY.md index load into every session. It is the Claude-authored counterpart to the CLAUDE.md you write. See Memory.

Auto mode

A permission mode in which a separate classifier model reviews Claude's actions instead of asking you about each one. Your explicit ask rules still prompt. It is designed to block scope creep, untrusted infrastructure and prompt injection, and it never sees tool results, so hostile text in a file cannot talk to it directly. From v2.1.283 it is the built-in starting mode for interactive terminal and VS Code sessions (on earlier versions, only on Pro, Max and Team plans). See Permission modes.

B

Bare mode

Starting with --bare skips hooks, skills, custom commands, subagents, installed plugins, MCP servers, auto memory and CLAUDE.md (skills in an --add-dir directory still load). It is what you want in CI or scripts where every machine should behave identically. See Headless.

Bundled skills

Skills that ship with Claude Code, such as /batch, /code-review, /debug and /loop. Unlike built-in commands, which run fixed logic, these hand Claude a detailed prompt and let it orchestrate the work, so they can spawn agents and adapt to your codebase. See Skills.

C

Channel

An MCP server that pushes events into a running session, so Claude can react to things happening while you are away from the keyboard. Channels can be two-way, replying back the way the message came. See Channels and the channels reference.

Checkpoint

The restore point recorded at the start of each turn you begin. /rewind (or Esc twice) uses checkpoints to restore code, conversation or both, or to summarise part of the conversation. They persist with the session but do not cover changes made through Bash, and they are not a substitute for git. See Checkpointing.

.claude directory

Where configuration lives: .claude/ at a project root for project settings, hooks, skills, subagents, rules and so on, and ~/.claude/ for your personal defaults. See The .claude directory.

CLAUDE.md

Your persistent instructions for Claude, loaded at the start of every session after the system prompt. Good for conventions, architecture notes and standing rules. It can live at ./CLAUDE.md or ./.claude/CLAUDE.md in a project, at ~/.claude/CLAUDE.md for you personally, or in managed policy for an organisation. All files found are concatenated, broadest first. The project-root file is re-read from disk after compaction. See Memory.

Cloud session

A session running on remote infrastructure (Anthropic-managed by default, or a self-hosted environment) that carries on after you close your laptop. Started from claude.ai/code, the mobile app, the desktop app with Cloud selected, claude --cloud, or a routine. Sessions on your own machine are local; reaching those from another device is Remote Control. See Claude Code on the web.

Command

Something you run by typing /name. Built-ins such as /clear, /model and /compact control the session; you can also add your own in .claude/commands/ or via plugins, though skills are now the preferred way to package them. Do not confuse these with claude CLI subcommands like claude mcp add, or with the command field of an MCP server config. See Commands.

Compaction

Automatic summarising of the conversation as the context window nears its limit: old tool output is cleared first, then the conversation is summarised. Project-root CLAUDE.md and auto memory reload from disk afterwards, but instructions you only gave in chat may not survive. /compact runs it on demand, optionally with a focus such as /compact keep the migration plan. See Context window.

Connector

An MCP server attached to your claude.ai account rather than configured locally. Sign in to Claude Code with that account and your connectors appear in /mcp. Organisations can provision connectors and control individual tools. See MCP.

Context window

The session's working memory: conversation, file contents, command output, CLAUDE.md, memory, loaded skills and system instructions. /context shows what is using it. See Context window.

D

Dispatch

Sending a coding task from the Claude mobile app that spawns a session in the desktop app on your computer. Pro and Max plans. See Desktop.

E

Effort level

How much the model is allowed to think on each step under adaptive reasoning. Higher effort is slower and more thorough; lower is quicker and cheaper. Supported on Fable models, Opus 4.6 and later, Sonnet 4.6 and later, and Haiku 5.5. See Model configuration.

Extended thinking

The model's visible reasoning before it answers, shown in grey italics in the terminal. Tune it with effort, or cap it with MAX_THINKING_TOKENS on models with a fixed thinking budget.

F

Frontmatter

The YAML block between two --- lines at the very top of a Markdown file. Skills, subagents, output styles and rules all read their settings from it (a skill's description, a subagent's tools) and treat the rest of the file as instructions. The opening --- must be the first line.

H

Hook

A handler you configure to run automatically at a fixed point in the lifecycle, such as before a tool runs, after an edit, or when a session starts. Handlers can be shell commands, HTTP endpoints, MCP tools, LLM prompts or subagents. Because they fire at fixed points rather than at the model's discretion, they are deterministic. Each hook has an event (when), a matcher (which occurrences) and a handler (what runs). See the hooks guide and hooks reference.

M

Managed settings

Organisation-wide settings that users and projects cannot override, delivered either from the admin console (server-managed) or as files and MDM policies on each device. Used for security policy, compliance and standard tooling. See Managed settings and Server-managed settings.

MCP

The Model Context Protocol, an open standard for plugging AI tools into external systems. Servers give Claude tools for things like Slack, Jira, databases or browsers. Manage them with /mcp or .mcp.json. See MCP.

MCP server

A program offering tools, prompts or resources over MCP. Add one with claude mcp add, in .mcp.json, via a plugin, or as a claude.ai connector. A stdio server is a process Claude Code launches from its command and args.

MCP Tool Search

Deferring MCP tool schemas until needed. Only names and server instructions load at start-up, and Claude fetches a tool's full schema when it decides to use it, so idle servers cost little context. See MCP.

N

Non-interactive mode

Running one prompt and exiting, with -p or --print. Used in CI, scripts and pipes. Still saved as a resumable session unless you pass --no-session-persistence. Formerly called headless mode; the Agent SDK is the programmatic equivalent. See Headless.

O

Output style

A set of instructions that changes how Claude responds (tone, format, behaviour). Unlike CLAUDE.md, which adds context alongside the default engineering instructions, a custom output style can replace those defaults. See Output styles.

P

Permission mode

The session's baseline approach to approvals: default, acceptEdits, plan, auto, dontAsk or bypassPermissions. Cycle with Shift+Tab in the terminal or use the mode picker elsewhere. default is labelled Manual in the CLI, IDE extensions and desktop app, and manual is accepted as an alias. See Permission modes.

Permission rule

A settings entry that allows, asks about or denies a tool call by tool name and argument pattern. Evaluated deny, then ask, then allow; the first match wins. Rules sit on top of the permission mode. See Permissions.

Plan mode

A permission mode where Claude investigates and proposes but does not edit source files. It can read, search and run exploratory commands, then presents a plan for approval. Enter with /plan or Shift+Tab.

Plugin

A packaged bundle of skills, hooks, subagents, MCP servers and more, installed as one unit and usually distributed through a marketplace. Plugin skills are namespaced plugin-name:skill-name. See Plugins.

Project trust

The dialog in which you accept a folder before Claude Code loads its repository-supplied configuration. Trust is remembered per folder, except your home directory, where it lasts for the session only. Until trusted, content such as project allow rules and project marketplaces is held back. See Permissions.

Prompt injection

Hostile instructions hidden in a file, web page or tool result that try to steer Claude somewhere you never asked. Defences include permissions, command injection detection and trust checks, and auto mode adds scanning of tool results and a classifier that never sees them. See Security.

R

Remote Control

Driving a session that runs on your own machine from your phone or a browser via claude.ai. Code and files stay local; only the interface is remote. See Remote Control.

Rules

Instruction files in .claude/rules/ that load alongside CLAUDE.md. Add paths: frontmatter to a rule and it only loads when Claude works with matching files, which keeps context lean. See Memory.

S

Sandboxing

Operating-system level filesystem and network isolation for the Bash tool. Inside the boundary you define, Claude can run commands without per-command prompts. It is a separate layer from permission rules. See Sandboxing.

Session

One conversation, tied to a working directory, with its own context window. Resume with claude -c, branch with --fork-session, or run several in parallel. /clear starts a new session while keeping the old one available through /resume. Transcripts live under ~/.claude/projects/. See Sessions.

Settings layers

Where configuration is read from, highest precedence first: managed policy, --settings flag, .claude/settings.local.json, .claude/settings.json, then ~/.claude/settings.json. Arrays merge across layers; single values from higher layers win. See Settings.

Skill

A SKILL.md file of instructions, knowledge or workflow that Claude loads when relevant or that you run as /skill-name. Skills follow the Agent Skills open standard, with Claude Code adding invocation control and subagent execution. They supersede custom commands: .claude/commands/ship.md and .claude/skills/ship/SKILL.md both give you /ship. See Skills.

Subagent

A helper with its own context window, system prompt, tool access and permissions, to which the main session delegates a task and from which it receives a summary. Good for keeping big searches out of your main context or running research in parallel. Built-ins include Explore, Plan and general-purpose. Subagents stay inside their session; to pass findings between sessions, see Cross-session messaging. See Subagents.

Surface

Anywhere you use Claude Code: terminal, VS Code, JetBrains, desktop app or claude.ai. They share one engine. Local sessions read your local configuration; cloud sessions start from a fresh clone and do not see your ~/.claude/. Slack and the Chrome extension connect to a surface rather than being one. See Platforms.

System prompt

The instructions sent ahead of the conversation on every request, covering tool use, safety and formatting. Add to it with --append-system-prompt or replace it with --system-prompt (see the CLI reference). It is the first layer of the prompt cache. CLAUDE.md and output style instructions are not part of it; they arrive as system reminders.

System reminder

Context Claude Code inserts into the conversation itself, rather than something you type. Your CLAUDE.md files, output style instructions, text returned by hooks as additionalContext, the list of available skills, notices that a file changed on disk, and commit attribution lines all arrive this way. In logged API requests they appear inside <system-reminder> tags in a user message, or on some models as a separate system message.

T

Teleport

/teleport pulls a cloud session into your local terminal, fetching its branch and conversation so you carry on from where it was. --cloud goes the other way. See Claude Code on the web.

Tool

An action Claude can take: read a file, edit code, run a command, search the web, start a subagent. Without tools Claude could only reply with text. See the tools reference.

Transcript

The saved file behind a session, by default ~/.claude/projects/<project>/<session-id>.jsonl. Resuming reads it back. The on-screen view is the transcript viewer described in Interactive mode.

Turn

One complete response from Claude: it starts when you send a message and ends when Claude finishes, with any number of tool calls between. Stop hooks fire at the end of each turn.

V

Verification loop

Giving Claude a check it can run (tests, a build, a screenshot comparison) and having it iterate until the check passes, rather than declaring victory after one attempt. It is the prerequisite for /goal, unattended runs and workflows, because without one the only judge of "done" is the agent itself. See Best practices and Goal.

W

Worktree isolation

Running Claude in a separate git worktree under .claude/worktrees/ (via -w, or isolation: worktree in a subagent definition) so its changes live on their own branch and folder, and parallel agents cannot trample each other. See Worktrees.

Old names you may still see

Related

• How Claude Code works
• Features overview
• Commands
• Tools reference