Cursor glossary
Explanation: what a concept is, why it exists, and what it is easy to confuse with. How to configure it: cheatsheet. How to click it: tutorial.
Map
┌──────────────┐
│ model+tools │
└──────┬───────┘
│
┌────────────┼────────────┐
│ │ │
┌────┴────┐ ┌───┴────┐ ┌────┴────┐
│ Rules │ │ Skills │ │ MCP │
│ AGENTS │ │Commands│ │ (ext.) │
└────┬────┘ └───┬────┘ └────┬────┘
│ │ │
└─────┬─────┴──────┬─────┘
│ │
┌─────┴─────┐ ┌────┴────┐
│ Agent │ │ Hooks │
│ Ask/Plan/ │ └─────────┘
│ Debug/Tab │
└─────┬─────┘
│
┌──────────┼──────────┐
│ │
┌─────┴─────┐ ┌──────┴──────┐
│ Subagents │ │ Cloud/Bugbot│
│ / CLI │ │ │
└───────────┘ └─────────────┘Rules / AGENTS.md are the standing brief. Skills / Commands are on-demand workflows. MCP turns external systems into tools. Hooks sit on the loop and do not rely on the model remembering. The Agent picks tools inside those constraints. Subagents move noisy work into another window. Bugbot and Cloud sit on the repo / PR axis.
Rules
What: Persistent instructions prepended to model context. Official wording: models do not keep memory across completions; rules supply reusable prompt-level context.
Why: Without them you restate “use pnpm” every thread. With them the team commits the convention.
Four triggers: Always / glob / description (“apply intelligently”) / @ only. Truth table: cheatsheet.
Versus AGENTS.md:
Rules (.mdc) | AGENTS.md | |
|---|---|---|
| Format | YAML frontmatter + Markdown | Plain Markdown |
| Trigger | Four modes | Present as project instructions |
| Best for | Path-scoped, needs metadata | Simple, portable, human-readable |
A plain .md under .cursor/rules is ignored. Official advice: use AGENTS.md if you want plain markdown.
Precedence: Team → Project → User. Merged; earlier source wins.
Docs: Rules
AGENTS.md
What: A normal Markdown file at the repo root (or a subdirectory) written for agents. No description / globs / alwaysApply.
Why: One file can serve Cursor, Claude Code, Codex, and anything else that reads AGENTS.md. Rules globs are stronger but Cursor-specific.
Role: Small repos can stop here. Larger ones: root AGENTS.md for globals, .cursor/rules/*.mdc per language or folder.
Docs: Rules · AGENTS.md
Skills
What: An Agent Skills directory: SKILL.md plus optional scripts/, references/, assets/. Discovered at startup. The agent uses description to decide relevance; you can also type /skill-name.
Why: Always-on rules consume the window every turn. Long runbooks should load on demand.
Versus Commands: Commands run when you type /. Skills may auto-apply unless disable-model-invocation: true. Cursor 2.4 ships /migrate-to-skills for dynamic rules and slash commands.
Built-ins (official table, not a product inventory): /create-rule, /create-skill, /create-subagent, /review-bugbot, /babysit, and others. Recheck Skills before documenting a slash name.
Docs: Skills
Commands
What: Markdown in .cursor/commands/, ~/.cursor/commands/, or Dashboard Team Commands. Listed when you type /.
Why: Prompts you run many times a day with a fixed script (open a PR, apply an issue template) without packing a Skill.
Status: Official page still says beta. Do not declare them removed.
Docs: Commands
MCP
What: Model Context Protocol. Transports: stdio, SSE, Streamable HTTP.
Why: Without it you paste Figma JSON and Sentry traces into chat. With it the agent pulls live systems.
Versus Hooks: MCP is a tool the agent can call. A hook is your script around a call. They stack: a hook can audit or block an MCP tool.
Safety: Servers run code and call APIs as you. Official notes: trusted sources, least-privilege keys, read the source for anything important.
Docs: MCP
Hooks
What: Processes spawned on Agent / Tab / workspace events. JSON over stdin/stdout. They can observe, block, or change behavior.
Why: Models will not reliably “always run prettier” or “never DELETE production.” Hooks keep policy outside the prompt.
Three families: Agent hooks, Tab hooks, workspaceOpen. Cloud agents run command-based project hooks only (no home-dir hooks, no prompt-based hooks). Reasons are in the official Hooks page (read-only early turns, no IDE session boundary, etc.).
Docs: Hooks
Subagents
What: Specialists the parent can spawn. Own context window. Foreground (blocking) or background (parallel).
Why: Official built-ins came from conversations that blew the window:
| Subagent | Job | Why isolated |
|---|---|---|
| Explore | Search and analyze | Huge intermediate output; faster model; many parallel searches |
| Bash | Shell series | Noisy logs |
| Browser | Browser MCP | DOM / screenshots |
Versus Skills (official table): isolation, parallelism, multi-step expertise → subagent. One-shot repeatable action → skill.
Cost: Each subagent burns its own tokens. Simple work is often faster on the parent.
Docs: Subagents
Modes
What: Constraints on one agent, not four products.
| Mode | Constraint |
|---|---|
| Agent | May edit and run commands |
| Ask | Read-only (changelog: search tools included; @Codebase tool gone) |
| Plan | Research, questions, editable plan; code after approval |
| Debug | Hypotheses → instrumentation → you reproduce → evidence |
Why: Default Agent wants to write code. Plan separates what from how. Debug stops unmotivated edits. Ask keeps an orientation thread from touching files.
Tab
What: A completion model. Accept (Tab) and reject (Esc) become signal for later suggestions.
Why: A full agent turn is the wrong tool for “change the next three lines.” Tab uses recent edits, lints, and accepted suggestions.
Not Chat. No tool loop. It does not treat your Rules folder as a task brief.
Docs: Tab
Bugbot
What: A service that reviews a PR diff (or the local diff /review-bugbot sees). It looks for bugs, security issues, and quality problems, then comments with fix ideas.
Why: Humans fatigue on large AI diffs. Bugbot reruns on each push and can follow .cursor/BUGBOT.md / team rules.
Not Debug Mode. Debug instruments your workspace. Bugbot reads a diff; it does not boot your app. The old overview matrix called Bugbot a “runtime debugger.” That was wrong.
Autofix: Official feature name. It spawns a Cloud Agent to patch findings. There is no separate 2026 product page titled Fixer.
Docs: Bugbot
Checkpoints
What: Automatic snapshots of files the agent is about to change. Preview and Restore from the timeline.
Why: Exploration and large refactors need an undo that is not “one more prompt.” Checkpoints are not git commits and not Plan files.
Docs: Agent Overview · Checkpoints
Codebase Indexing
What: The repo is chunked at function/class granularity, embedded, and searched by meaning. “Where is the top nav?” can hit header.tsx.
Why: grep misses synonyms. Official claim: embeddings are computed at index time, so agent search is faster and cheaper.
Privacy (official): paths encrypted; source not stored in plaintext long-term; processed in memory while indexing. Defaults exclude .gitignore / .cursorignore. Semantic search is available around 80% index completion. Idle indexes are deleted after about six weeks.
Docs: Semantic search
Privacy Mode
What: A governance switch. When on, Cursor and model providers do not train on your code. On by default for Enterprise teams.
Why: Two exits exist: LLM requests, and (optional) encrypted repo copies for Cloud Agents. Privacy Mode is about training / retention promises. It is not “code never leaves the machine.”
Cloud Agents: Officially the only feature that requires Cursor to store code. If policy forbids storage, leave Cloud Agents off. Everything else still works.
Docs: Privacy and Data Governance
Cloud Agents
What: Agents on isolated VMs. They clone, branch, work, and open PRs. Start from the editor Cloud dropdown, cursor.com/agents, mobile, Slack / GitHub / Linear @cursor, or a CLI message prefixed with &.
Former name: Background Agents (official Naming History). Do not document them as two products.
Why: Work that should not occupy the local window: drive-by bugs, tests, docs, many parallel runs. Cousins: Claude remote / Dispatch, Gemini Jules.
Versus local subagents: Local children use your machine and local MCP. Cloud uses team MCP and the cloud environment.
Versus Bugbot: Cloud edits a repo. Bugbot reviews a PR diff. Autofix is the bridge.
Docs: Cloud Agent, Setup, agent-best-practices. Tutorial: Cloud Agents.
Cursor CLI
What: Official terminal entry. The binary is agent. Interactive session or agent -p for headless / CI. Same Agent / Ask / Plan modes. Loads .cursor/rules, AGENTS.md, and CLAUDE.md.
Why: You are already in SSH / tmux / CI and should not open a GUI just to reuse the same rules. This is another surface of the same agent, not a third model.
Versus editor / Cloud: The editor has Tab and Inline Edit; the CLI does not. Cloud runs on an isolated VM; the CLI edits the current working tree unless you pass --worktree. Prefix a message with & to hand off to Cloud.
Docs: CLI Overview, Installation, Headless. Tutorial: Cursor CLI.
Design Mode
What: Visual prompting in the browser inside the Agents Window. Click an element, multi-select, draw on a frozen viewport frame, or speak. Toggle Cmd+Shift+D.
Why: UI changes are spatial. The agent gets element identity (xpath, component, computed styles, fiber props) plus a screenshot, not just a sentence.
Versus Claude Design: This is an overlay on your running app. Claude Design is a standalone brand-aware canvas.
Docs: Design Mode. Tutorial: Design Mode.
Origin
What: Cursor's git forge (early beta). Host repos, mirror GitHub, browse / search at cursor.com/codebase, open and merge PRs. The CLI binary is origin, not agent.
Why: You want Cursor-hosted storage and PRs, not only review comments on GitHub.
Not Rules. Claiming a codebase name enables it. Pro / Teams / Enterprise only. Follows the namespace owner's Privacy Mode. Legacy privacy cannot enable it.
Docs: Origin. Tutorial: Origin.
Security Agents
What: Two Cursor-managed agents on Automations: Security Reviewer (PR / MR) and Vulnerability Scanner (cron). They require Cloud Agents. Local skill: /review-security (Cursor 3.7+; CLI coming soon).
Why: Dedicated security gate and at-rest scan, billed to the team usage pool under a shared service account.
Versus Bugbot: Bugbot comments on a general PR diff. Security Agents are a separate managed family. PR Routing can read both.
Docs: Security Agents. Tutorial: Security Agents.
PR Routing & Approval
What: Automations agent that assigns reviewers from ownership / history and can approve low-risk PRs. Exact policy file: APPROVAL_POLICY.md. Optional routing: .cursor/approval-policies/ROUTING.md.
Why: Route the right humans; auto-approve only where policy allows. It does not replace a full code review.
Gotcha: Wrong filename is ignored. A PR that edits its own policy cannot use the new text to relax that same PR.
Docs: PR Routing & Approval. Tutorial: PR Routing.
Cursor SDK
What: Call the same agent from your process. TypeScript @cursor/sdk (Node 22.13+), Python cursor-sdk (3.10+), other languages via the SDK Bridge. Runtime is local or cloud; local is the loop + filesystem, not a local model.
Why: CI bots, triage workers, in-product agents, orchestrators. Same pricing / Privacy Mode as IDE and Cloud; spend tagged SDK.
Keys: user or service-account. Team Admin API keys are not supported yet.
Docs: TypeScript, Python, Bridge. Tutorial: Cursor SDK.