Glossary of AI Development Terms

You are reading a tutorial and hit a term you have never seen before — AGENTS.md, worktrees, sub-agents, thinking tokens. This glossary defines every piece of jargon you will encounter across Cursor, Claude Code, and Codex so you can get back to shipping code.

A

Agent Mode

A Cursor feature where the AI can autonomously read, modify, and create files across your project. Unlike Ask mode (read-only), Agent mode executes multi-step tasks: editing code, running terminal commands, fixing lint errors, and iterating until the job is done.

Agent Skills

An open standard for extending AI coding agents with reusable, shareable capabilities. Installed via npx skills add <owner/repo> from the Skills.sh marketplace. Skills work across Claude Code, Cursor, Codex, and 35+ other agents. Lighter-weight than MCP servers — they augment a single capability rather than maintaining a persistent connection.

AGENTS.md

A project-level configuration file recognized by Codex (and increasingly by other tools) that defines how AI agents should behave in a repository. Similar in purpose to CLAUDE.md but follows the AGENTS.md convention. Placed at the project root or in subdirectories. Specifies coding standards, test commands, deployment patterns, and tool preferences.

API Key

A secret string used to authenticate requests to an AI provider (Anthropic, OpenAI, Google). Can be personal or organizational. Store securely — never commit to version control.

Ask Mode

Cursor’s read-only conversation mode. The AI can read your codebase and answer questions but cannot modify files. Use it for exploration, planning, and understanding before switching to Agent mode for execution.

@ Symbols

Context references in Cursor and Claude Code that let you include specific files, folders, URLs, or documentation in your AI conversation. Examples: @file.js, @docs, @codebase, @web. Codex uses similar file-referencing patterns in its prompt interface.

Automations

Codex feature that lets you trigger agent runs automatically in response to events — GitHub issues, Slack messages, Linear tickets, or scheduled cron jobs. Runs in the cloud without manual intervention.

B

Background Agent

An agent that runs asynchronously while you continue working. In Cursor, triggered via Cmd/Ctrl+E or from Slack. In Codex, cloud tasks run as background agents by default. The agent completes the task independently and notifies you when done.

BugBot

Cursor’s automated pull request reviewer. Analyzes PRs for bugs, security issues, and code quality problems. Integrates with GitHub and posts inline comments.

C

Checkpoint

A saved state of your working directory created by Cursor before applying AI-generated changes. If the changes break something, you can revert to the checkpoint instantly. Think of it as an automatic git stash for every AI interaction.

CLAUDE.md

A markdown file that Claude Code reads automatically for project-specific context. Place it at the project root (or in subdirectories for scoped rules). Defines coding standards, common commands, architecture notes, and any instructions the AI should always follow. Created with /init and updated with /memory or the # shortcut.

Cloud Agent

An agent that runs on remote infrastructure rather than your local machine. Codex Cloud agents execute in sandboxed environments with their own compute resources. Cursor’s Background Agent can also hand off tasks to the cloud using the & suffix.

Codebase Indexing

The process of analyzing your entire project to enable semantic search. Cursor creates vector embeddings of your code. Claude Code uses agentic search (reading files on demand). Codex indexes repositories when you connect them.

Codex

OpenAI’s multi-surface AI coding agent. Available in the ChatGPT desktop app, as a CLI tool, an IDE extension, and a Cloud service. It offers GPT-5.6 Sol, Terra, and Luna according to plan and workload. Isolated desktop tasks can use worktrees; other surfaces use their own working-directory or cloud isolation model.

Compact / Compaction

The process of summarizing conversation history to free up context window space. In Claude Code, triggered with /compact. In Codex, compaction happens automatically when the context approaches the model’s limit. Preserves key decisions while discarding verbose intermediate steps.

Composer

Cursor’s multi-file editing interface (now largely merged into Agent mode). Describes changes across multiple files in a unified view before applying them.

config.toml

Codex’s configuration file, typically located at ~/.codex/config.toml. Defines the default model, approval policy, sandbox, and MCP/provider behavior. Credentials are handled by codex login or environment/secret management, not stored in this file.

Context Window

The maximum amount of text (measured in tokens) that a model can process in one request. Current Claude Fable 5, Opus 4.8, and Sonnet 5 tiers support 1M tokens; GPT-5.6 Sol, Terra, and Luna expose 1.05M through the API; Gemini 3.1 Pro supports 1M. Host products can expose smaller effective windows. Managing context effectively is one of the most important skills in AI-assisted development.

Custom Commands

User-defined slash commands. In Claude Code, created as markdown files in .claude/commands/. In Codex, defined through configuration. They let you package frequently used prompts into reusable /my-command shortcuts.

D

Deep Reasoning

Reasoning controls that allocate more computation to complex problems. In current Claude Code, use the thinking toggle and effort level (/model or /effort) for durable control. think / think hard are ordinary prompt text; ultrathink is a special one-turn deeper-reasoning hint that does not change API effort. Codex and Cursor expose their own model- and surface-specific controls.

Diff View

A visual representation of changes between original code and AI-suggested modifications. All three tools provide inline diff views for reviewing changes before accepting them.

E

Effort Parameter

A setting on Claude models (Fable 5, Opus 4.8, Sonnet 5) that controls how much reasoning computation the model uses. Lower effort = faster and cheaper; higher effort = deeper analysis. Fable 5 supports five levels: low, medium, high, xhigh, and max. Adjustable via API or through tool interfaces.

Embeddings

Mathematical vector representations of code that enable semantic search. Cursor creates embeddings of your entire codebase so you can search by meaning, not just keywords. Claude Code and Codex use different retrieval strategies but achieve similar results.

Enterprise Mode

Features designed for corporate environments: plan-specific privacy and retention controls, SSO/SAML authentication, audit logging, SCIM provisioning, and applicable compliance certifications.

F

Fable 5

Anthropic’s most capable model (released June 9, 2026) — the new tier above Opus 4.8. Model ID claude-fable-5; $10/$50 per million tokens; 1M token context window, 128K max output. Excels at complex multi-file refactorings, bug-fixing, building apps from scratch, and long-running tasks. Available in Claude Code (v2.1.170+, /model fable), Cursor’s model picker, and the Claude API.

Fast Requests (Historical)

A legacy Cursor quota term. Current plans meter agent usage against plan allowances and model pricing rather than a documented Fast/Slow request split; check Settings > Usage for the live model.

Full-Auto Mode (Deprecated)

The old Codex --full-auto compatibility shortcut. Current configuration treats approval and sandbox as separate controls: approval_policy = "never" suppresses prompts but does not expand the sandbox. Use never only for trusted unattended work with an explicitly chosen sandbox.

Fusion Model

Cursor’s specialized Tab completion model capable of multi-line predictions and long-range code jumps across files.

G

Git Integration

Built-in version control capabilities that let AI agents create branches, write commits, resolve merge conflicts, and open pull requests. All three tools integrate deeply with Git and GitHub.

Git Worktrees

A Git feature that lets you check out multiple branches simultaneously in separate directories. Codex can use worktrees for isolated tasks in the ChatGPT desktop app, preventing conflicts when running parallel local tasks; CLI and Cloud use different execution models.

H

Headless Mode

Running an AI agent non-interactively, typically in CI/CD pipelines or automation scripts. Claude Code supports this with claude -p "task". Codex supports it through its Cloud API and CLI flags. No human in the loop — the agent executes and exits.

Hooks

Automation triggers at specific points in an agent’s workflow. In Claude Code, hooks run before or after specific tool calls (file edits, command execution). Defined in .claude/settings.json. Used for linting, formatting, validation, or custom logic.

I

IDE (Integrated Development Environment)

A software application for writing code. Cursor is an AI-native IDE (forked from VS Code). Codex offers an IDE extension for supported VS Code-compatible environments. Claude Code is terminal-native and also has integrated IDE extensions.

Inline Edit

Cursor’s Cmd/Ctrl+K feature for making targeted AI edits to selected code directly in the editor, without opening a chat panel.

Intent Context

The “what you want to achieve” part of a prompt (prescriptive), as opposed to state context which describes the current situation (descriptive). Good prompts balance both: describe the current state, then clearly state the desired outcome.

L

Large Language Model (LLM)

An AI system trained on vast text datasets. Current examples in these tools include Claude Fable 5 / Opus 4.8 / Sonnet 5, GPT-5.6 Sol / Terra / Luna, Grok 4.5, Composer 2.5, and Gemini 3.1 Pro.

Long Context Mode

A mode for processing large amounts of text that pushes against context window limits. In Cursor, this is Max mode (token-based pricing). In Claude Code, the model’s 1M window handles most cases natively.

M

Max Mode

Cursor’s enhanced context mode that uses token-based pricing to provide maximum model capabilities and the full context window for complex tasks.

MCP (Model Context Protocol)

An open standard created by Anthropic for connecting AI assistants to external tools and data sources. MCP servers expose capabilities (database queries, API calls, browser automation, file operations) that any MCP-compatible agent can use. All three tools — Cursor, Claude Code, and Codex — support MCP.

MCP Server

A program that implements the MCP specification and exposes one or more tools to AI agents. Examples: GitHub MCP (PR and issue management), Postgres MCP (database queries), Puppeteer MCP (browser automation), Context7 MCP (library documentation).

Memory System

Persistent project-specific knowledge that survives between sessions. Cursor uses the “Memories” feature. Claude Code uses CLAUDE.md files and the /memory command. Codex uses AGENTS.md and its project configuration.

Message Limit

Plan capacity consumed by interactions, tokens, context, tools, model, and effort. Claude Code Pro and Max are usage-limited, but a fixed message estimate is unreliable; Codex and Cursor likewise meter usage according to plan and workload.

Multi-file Editing

The ability to coordinate changes across multiple files in a single operation. A core strength of all three tools — Agent mode in Cursor, default behavior in Claude Code, and standard operation in Codex.

N

Notifier Integrations

Codex’s ability to send notifications and receive triggers from external services. Supports GitHub (issues, PRs), Slack (messages, threads), and Linear (tickets). Enables automated workflows where an agent responds to real-world events.

O

Opus 4.8

Anthropic’s Opus-tier model one tier below Fable 5. Released May 2026. Strong agentic performance and reasoning. It is the account default on Max, Team Premium, Enterprise pay-as-you-go, and Anthropic API sessions, while other accounts/providers can default differently. Available in Claude Code, Cursor, and via API.

P

PAT (Personal Access Token)

A secure authentication token for Git services like GitHub. Required for CLI operations and many MCP server integrations.

Permission Mode

Controls when an AI agent asks before an action. Claude Code offers granular permission settings per tool. In Codex, on-request is the interactive policy and never suppresses prompts; sandbox policy separately controls what can execute. The old --full-auto shortcut is deprecated.

Planning Mode

An explicit phase where the AI breaks down a complex task into structured steps before implementing anything. In Claude Code, use Plan Mode (Shift+Tab or /plan) and adjust /effort when needed; think is ordinary prompt text. Cursor and Codex expose planning through their current agent interfaces.

PRD (Product Requirements Document)

A specification of features and functionality. The PRD-to-Plan-to-Todo methodology is a proven pattern for complex features across all three tools: write the PRD, ask the AI to create a plan, convert the plan into a checklist, then execute.

Privacy Mode

A product/plan setting and contractual control governing training, retention, and data handling. Critical for enterprise compliance and IP protection; verify the exact business or consumer terms for each tool.

Prompt Engineering

The practice of crafting effective instructions for AI models. Includes techniques like few-shot examples, chain-of-thought reasoning, explicit constraints, and structured output formats. The single highest-leverage skill for AI-assisted development.

R

Rules Directory

Cursor’s .cursor/rules/ folder containing project-specific AI behavior guidelines in markdown files. Rules are automatically loaded based on file context, glob patterns, or manual selection. Equivalent in function to CLAUDE.md and AGENTS.md.

S

Sandbox

An isolated execution environment that constrains what AI-generated commands can affect. Codex sandbox policy is surface-specific and separate from approval; isolated ChatGPT desktop tasks may also use worktrees. Claude Code and Cursor provide their own local and managed isolation controls.

Semantic Search

AI-powered search that understands meaning rather than just matching keywords. Cursor’s @codebase search uses embeddings. Claude Code uses agentic search (reading and reasoning about files). Both find conceptually relevant code even when the exact terms differ.

Slash Commands

Special commands prefixed with / that trigger built-in or custom actions. Claude Code includes /help, /clear, /compact, /model, /effort, and /simplify (distinct from /code-review); there is no /think command. Codex: /model, /new, /review, /status (type / to see the live list). Cursor commands depend on the current agent surface.

Slow Requests (Historical)

The former fallback half of Cursor’s Fast/Slow quota model. It is not the current documented plan structure; consult Settings > Usage and Cursor’s live pricing page.

SOC 2

A security compliance framework for service controls. Availability and scope of SOC reports depend on vendor, product, and contract; verify the current trust center rather than inferring coverage from a plan name.

Sonnet 5

Anthropic’s cost-effective model with a 1M token context window and 128K max output. Its introductory $2/$10 per-MTok rate through August 31, 2026 is 40% of Opus 4.8’s $5/$25 rate. It is the account default on Pro, Team Standard, and Enterprise subscription seats; other plans/providers differ.

State Context

Information describing the current state of code, errors, or environment (descriptive). Combined with intent context (what you want to achieve), it forms a complete prompt.

Sub-agent

A separate AI instance spawned by the main agent to handle a specific subtask. In Claude Code, sub-agents run with their own context while the parent agent continues. In Cursor, subagents can be specialized for testing, documentation, or refactoring (each runs in its own context window). In Codex, parallel cloud tasks function similarly.

T

Tab Autocomplete

Cursor’s predictive code completion feature. Suggests multi-line completions based on surrounding context. Accept with Tab, accept partially with Cmd/Ctrl+Right Arrow, reject with Escape.

Terminal Integration

Direct command-line access within the tool. Claude Code is terminal-native. Cursor integrates terminals in the IDE. Codex CLI runs directly in your shell.

Thinking Tokens

The computation allocated to model reasoning. Current Claude Code uses a thinking toggle and effort levels rather than /think or /ultrathink slash commands. The ultrathink prompt hint affects one turn in context without changing API effort; higher persistent effort can improve hard-task quality while increasing usage and latency.

Token

The basic unit of text processed by LLMs. Roughly 3-4 characters in English. Context windows, pricing, and rate limits are all measured in tokens.

Tool Use

The ability of an AI model to call external functions — reading files, executing commands, making API requests, querying databases. The foundation of agentic behavior. MCP standardizes how tools are exposed to models.

U

Ultrathink

A one-turn prompt hint recognized by Claude Code that adds an in-context instruction for deeper reasoning. It does not change the API effort level or reserve a fixed token budget; use the thinking toggle or /effort for durable control.

V

Version Control Integration

Built-in Git capabilities: commits, branches, diffs, pull requests, merge conflict resolution. All three tools can operate Git autonomously when permitted.

Vim Mode

An optional keyboard input mode in Claude Code that enables vim-style navigation and editing in the prompt input. Configure Editor mode through /config; the old /vim command was removed.

W

Worktree

A Git worktree is a separate working directory linked to the same repository. Codex can create one for an isolated task in the ChatGPT desktop app so parallel local tasks do not interfere; this is not universal across CLI, IDE, and Cloud. See also: Git Worktrees.

Workspace

The root directory (and subdirectories) of your current project. All three tools scope their operations to the workspace.

Y

YOLO Mode

An informal term for suppressing permission prompts. In Claude Code this can mean --dangerously-skip-permissions. In Codex, use approval_policy = "never" only for trusted unattended work and choose the sandbox separately; --full-auto is deprecated. Cursor permissions depend on the current agent surface. Use version control and strong isolation.

Extensibility: Agent Skills | MCP | MCP Server | Skills.sh

Security and Enterprise: Enterprise Mode | Privacy Mode | SOC 2 | Permission Mode