๐ Glossary
A reference for terms that come up frequently in Claude Code and the AI development ecosystem.
Aโ
Agentโ
An AI system that acts autonomously to achieve a specific goal. Beyond simply answering questions, an agent uses tools and works through multiple steps to complete a task. Claude Code itself is an agent, and you can build custom agents with the Agent SDK.
Agent SDKโ
A toolkit for running Claude Code programmatically. It offers three interfaces: the CLI (claude -p), a Python SDK, and a TypeScript SDK. It leverages Claude Code's agentic loop to build automation pipelines.
Agent Teamโ
An experimental feature where multiple Claude Code instances collaborate on work simultaneously. Configure the display mode with --teammate-mode, and detect idle states with the TeammateIdle hook.
Agentic Loopโ
The iterative cycle in which an agent calls tools, receives results, and decides its next action. The maximum number of iterations is controlled with maxTurns.
allowedToolsโ
One of the Agent SDK query() options. Specifies the list of tools the agent may use as an array. (Example: ["Read", "Write", "Bash"])
Anthropicโ
The AI safety research company that developed Claude. It offers products including Claude Code, the Claude API, and Claude.ai.
Auto memoryโ
Notes that Claude saves for itself based on the user's corrections and preferences. Stored per git repository under ~/.claude/projects/. All worktrees of the same repository share one directory. The first 200 lines (or 25KB) of the MEMORY.md index are loaded automatically at the start of every session. It is the Claude-written counterpart to CLAUDE.md (which the user writes).
Bโ
Bash (tool)โ
A built-in tool that lets Claude Code execute shell commands. Used when Claude runs terminal commands like npm test or git log directly.
Bare modeโ
The --bare startup flag. Skips automatic discovery of hooks, skills, plugins, MCP servers, auto memory, and CLAUDE.md, applying only explicit flags. Recommended for CI and scripted invocations where you need identical behavior regardless of the machine.
budget_tokensโ
The maximum number of tokens Claude can spend on internal thinking in Extended Thinking. Higher values allow deeper reasoning but also increase cost.
Cโ
Channelโ
A type of MCP server that pushes events that occurred while you were away into an ongoing session. Two-way flow is also possible: Claude reads events arriving in the inbox and responds through the same channel. Telegram, Discord, and iMessage are included in the research preview.
Checkpointโ
A restore point automatically created for every prompt you send. Claude Code saves a file snapshot right before each edit, so you can roll back to a checkpoint. Press Esc twice or run /rewind to restore code, conversation, or both to an earlier point, or to summarize the conversation from a selected message onward. Checkpoints are session-local and separate from git. Changes made through the Bash tool are not tracked.
CLAUDE.mdโ
A configuration file that Claude Code reads automatically at project startup. Record project rules, coding conventions, frequently used commands, and current progress here to give Claude context.
Claude Codeโ
A terminal-based AI coding assistant developed by Anthropic. It autonomously performs tasks such as reading/writing files, executing shell commands, and Git operations.
Context Windowโ
The maximum number of tokens a model can process at once. It includes conversation history, file contents, and the system prompt. When exceeded, the oldest content is truncated first.
/compactโ
A slash command used when the context window fills up. It summarizes the conversation so far to compress the context. For automatic compaction, the trigger threshold is adjusted with the CLAUDE_AUTOCOMPACT_PCT_OVERRIDE environment variable (default ~95%).
Compactionโ
The behavior of automatically summarizing the conversation as the context window approaches its limit. Older tool outputs are cleaned up first, then the conversation body is summarized. The project-root CLAUDE.md and auto memory survive compaction and are automatically reloaded, but instructions given only within the conversation can be lost. Manual invocation = /compact; to specify a focus = /compact focus on the API changes.
Dโ
Desktop Appโ
A way to run Claude Code in a desktop application instead of the terminal. During a conversation, the /desktop command moves the current session to the desktop app. Supported on macOS and Windows.
Dispatchโ
A phone-initiated task router: send a coding task from your phone and the Desktop app automatically creates a Claude Code session. Send a prompt from the Claude mobile app, and the Desktop app on your own machine receives it and performs the work. Supported on Pro and Max plans.
Eโ
Escalationโ
The process by which an agent detects a situation it cannot handle or one that is high-risk, and notifies a human or a higher-level system. An important safety mechanism in enterprise AI team design.
Edit (tool)โ
A built-in Claude Code tool that modifies a specific part of a file. Unlike Write, which overwrites the entire file, it performs partial edits via exact string matching.
Effort Levelโ
A setting that controls the depth of adaptive reasoning (dynamic reasoning based on task complexity) in Claude Code. Fable 5, Opus 4.8, and 4.7 support five levels โ low, medium, high, xhigh, max โ while Opus 4.6 and Sonnet 4.6 support low, medium, high, max. Adjust it via the /effort slider, the /model menu, or the CLAUDE_CODE_EFFORT_LEVEL environment variable. Defaults: Fable 5 and Opus 4.8 = high, Opus 4.7 = xhigh. Separately, ultracode is not an effort level but a Claude Code setting that bundles xhigh with dynamic workflow orchestration.
Extended Thinkingโ
A feature where Claude performs internal reasoning as separate tokens before responding. In Claude Code it is controlled via Effort Level. When using the API directly, specify the thinking budget with thinking.budget_tokens. Note that Fable 5 is adaptive-thinking only, so thinking cannot be turned off (thinking: disabled is not supported), and the raw thinking process is not returned.
Fโ
Fast Modeโ
A high-speed mode toggled with /fast in Claude Code. It uses the same Opus model at higher speed and is billed separately. In Opus 4.8, the fast mode unit price is lower. It is Opus-only, so Fable 5 does not support it; enabling it on another model automatically switches to Opus. Well suited for live debugging and rapid iteration.
Few-shot Promptingโ
A prompting technique that provides the model with examples (example input + example output) to guide it toward the desired output format.
Gโ
Glob (tool)โ
A built-in Claude Code tool that searches for files via file-pattern matching. Supports wildcard patterns like **/*.ts.
Grep (tool)โ
A built-in Claude Code tool that searches file contents with regular expressions. Used to find specific functions or patterns in a codebase.
Hโ
Headless mode โ Non-interactive mode (rename)โ
The official docs renamed this to "Non-interactive mode." A mode that runs Claude Code programmatically without a UI, mainly used in CI/CD pipelines and automation scripts. Run it with claude -p "prompt" or the --print flag. Behavior and flags are unchanged.
Hooksโ
Handlers that run automatically when specific events occur (before/after tool execution, response completion, and so on). Four types are supported: shell commands, HTTP endpoints, LLM prompts, and MCP tools. There are 18 events in total, used for notifications, logging, automated validation, permission control, and more.
Mโ
maxTurnsโ
An Agent SDK query() option. Limits the maximum number of iterations of the agentic loop. Important for preventing infinite loops and controlling cost.
MCP (Model Context Protocol)โ
An open standard protocol led by Anthropic. It connects external systems (databases, APIs, file systems, etc.) to AI models in a standardized way.
MCP Serverโ
A server that implements the MCP protocol. You can install MCP servers for services like Slack, GitHub, and Notion so Claude Code can integrate with them directly.
MCP Tool Searchโ
A context-saving mechanism that defers MCP tool schemas until they are needed. Only tool names are loaded at startup; when Claude decides to actually use a tool, it fetches the full schema. This keeps unused MCP servers from consuming large amounts of context.
Multi-agentโ
An architecture in which multiple agents collaborate on a single complex task. An orchestrator distributes the work, and specialized agents each handle their own role.
Mythos-classโ
The Claude model tier positioned above Opus. The generally available model is Fable 5 (released June 2026, Claude Code v2.1.170+), while Mythos 5, which has some safeguards removed, is limited to Project Glasswing partners. When Fable 5's safety classifier detects cybersecurity or bio/chemistry-related requests, Opus handles that request instead and a notice appears in the transcript โ the session then stays on that Opus model, and to return to Fable 5 you must run /model fable yourself (automatic switching can be toggled in /config). Access was temporarily suspended for all customers on 2026-06-12 under a US government export control directive, and was restored as of 2026-07-01 (official announcement: anthropic.com/news/redeploying-fable-5). For Pro, Max, Team, and some Enterprise plans, it is included in the subscription up to 50% of the weekly usage limit through July 7, after which it switches to usage credits.
Oโ
Output Styleโ
A setting that changes Claude Code's response style. There are three: Default, Explanatory (adds explanations about implementation choices and codebase patterns), and Learning (a hands-on mode where you write code yourself). Switch with /output-style.
Orchestratorโ
In a multi-agent system, the higher-level agent that analyzes work and distributes it to the appropriate agents. In this guide's capstone, the CEO agent plays the orchestrator role.
Pโ
Permission Modeโ
Sets the level of user approval Claude Code requires for each action, such as modifying files or running commands. There are four modes โ Default, Auto-accept Edits, Plan, and Bypass Permissions โ switched with Shift+Tab. Additional modes per the latest official docs: dontAsk and auto (Auto mode; Max, Team, Enterprise, and API).
Plan modeโ
A permission mode in which Claude investigates and proposes changes without modifying source files. It runs read, search, and exploration commands, but does not touch source until you approve. Enter it with /plan or Shift+Tab.
Pluginโ
A package that extends Claude Code's functionality. It can include Skills, Agents, Hooks, and MCP servers, and is shared/distributed through the marketplace. Defined by a .claude-plugin/plugin.json manifest.
Prompt Cachingโ
A feature that caches long, repeatedly used system prompts or documents on Anthropic's servers to reduce cost and latency. Configured with the cache_control: { type: "ephemeral" } marker.
Qโ
query()โ
The core function of the Agent SDK. It sends a task to Claude, runs the agentic loop, and returns the result.
Rโ
RAG (Retrieval-Augmented Generation)โ
A technique that retrieves relevant information from an external database and adds it to the model's response. It compensates for the model's knowledge limits and lets it use up-to-date information.
Rate Limitโ
A cap on API request frequency. Exceeding the allowed number of requests per minute/hour returns a 429 error. Queuing logic is needed when implementing multi-agent systems.
Read (tool)โ
A built-in Claude Code tool that reads file contents. Used for code analysis, checking documentation, and more.
Sโ
Sandboxโ
A security feature that isolates Claude Code's file system access. Toggled with /sandbox, and available only on supported platforms.
Skillโ
A SKILL.md file containing instructions, knowledge, and workflows. Claude loads it automatically at the relevant moment, or the user invokes it directly with /skill-name. It follows the Agent Skills open standard, which Claude Code extends with invocation control and subagent execution. It is the recommended successor to custom commands (.claude/commands/deploy.md), and both it and .claude/skills/deploy/SKILL.md work as /deploy. Existing commands/ files continue to work.
Streamingโ
Receiving a response in real time as it is generated rather than all at once. Useful for displaying long responses on screen immediately instead of waiting.
Subagentโ
A lower-level agent that performs a specific task under the orchestrator's direction. The unit of execution in a multi-agent architecture.
System Promptโ
A predefined instruction that specifies the model's role, behavior, and constraints. Set before the conversation begins and applied throughout the entire conversation.
Tโ
Tokenโ
The basic unit in which an AI model processes text. Korean has lower token efficiency than English, so the same content consumes more tokens. API pricing is based on token count.
Tool Useโ
The capability for Claude to call external tools (functions) to obtain information or perform actions. In the Claude API, tools are defined with the tools parameter.
Wโ
Worktreeโ
A feature that runs Claude Code on an isolated branch using Git worktrees. Used like claude -w feature-auth, and useful for running parallel Claude Code sessions simultaneously.
Write (tool)โ
A built-in Claude Code tool that creates or overwrites files. Used for creating new files or fully replacing existing ones. For partial edits, use the Edit tool.