Cursor Basics

What You Are Actually Installing

Cursor is an editor whose AI is the point rather than an extension bolted onto one. That framing matters on day one, because most of what surprises new users comes from treating it as "VS Code with a chat panel."

The install itself is ordinary: download, run the installer, open the app. On Linux the documentation sends you to your package manager first — apt or dnf — and falls back to extracting the AppImage or archive when your distribution has no package.

The Two Migration Surprises

If you are coming from VS Code, one fact will cost you an afternoon if you do not know it: Cursor uses the Open VSX extension registry, not the VS Code Marketplace. Cursor is a VS Code fork, so the assumption that every extension is one search away is reasonable and wrong. Some Marketplace-only extensions exist here as first-party builds instead — some Microsoft Marketplace-only extensions are replaced by audited Anysphere builds.

Your keybindings do come with you. Cursor uses the same default shortcuts as VS Code, plus shortcuts for AI features.

From JetBrains the shock is structural rather than cosmetic. Cursor uses a file-and-folder project model instead of JetBrains' project system — you open a folder, there is no project wizard — and language support comes from extensions rather than built-in plugins. An empty Python or Go experience is a missing extension, not a broken editor.

Rules Are Prompt Text, Not Configuration

Rules are how you give Cursor standing instructions, and almost every rule that "does not work" fails for one of four structural reasons rather than because the model ignored it.

First, rules are included at the start of the model context when they apply. They are prompt text, so length is a running cost on every request, not a free lookup.

Second, whether a rule applies at all is decided by its frontmatter. If alwaysApply is true, the rule will be applied to every chat session. Otherwise, the description of the rule will be presented to the Cursor Agent to decide if it should be applied. A rule set to apply intelligently with an empty description can never load — there is nothing to judge relevance against. A rule with no globs and no description is reachable only by mentioning it in chat.

Third, precedence runs in the opposite direction to most config systems. When rules conflict, precedence is: Team Rules > Project Rules > User Rules. Your personal preference loses to the repository, which loses to the team. And the sources do not replace one another — all applicable rules are merged; earlier sources take precedence when guidance conflicts.

Fourth, rules have a narrower reach than their name suggests. Rules do not impact Cursor Tab or other AI features, and User Rules are not applied to Inline Edit. A naming convention you want enforced in completions needs a linter, not a rule.

For simple cases skip the machinery: Cursor supports AGENTS.md in the project root and subdirectories, so a monorepo can scope instructions per package instead of paying for every convention in every conversation.

What The Agent Can And Cannot See

Two ignore mechanisms decide what reaches the model. Files ignored by git are also ignored by Cursor's indexing, so you rarely need to restate dependency folders, and an ignore rule you forgot is a common reason the agent insists a file does not exist.

The part worth memorising is the limit. Terminal commands and MCP tools run outside of Cursor's file access controls, so they may still be able to read ignored files. An ignore entry is a context-hygiene tool, not a security boundary. If a file must never reach a model, ignoring it is not sufficient.

Skills, MCP, And The Trust You Extend

Skills are the unit for a repeatable procedure. When Cursor starts, it automatically discovers skills from skill directories and makes them available to Agent — discovery is at startup, so a new skill appearing requires a restart, and the agent is presented with available skills and decides when they are relevant based on context. The description selects the skill; a vague one means it never fires. The dividing line against rules is worth stating plainly: use a rule when a short instruction is enough. Use a skill when Agent needs a detailed, repeatable process to follow.

MCP servers extend Cursor with outside tools. Configuration merges, and locality wins: both files are merged. If the same server name appears in both, the project-level config takes priority. After editing by hand you must save the file and restart Cursor. One server failing does not take the rest down — Cursor isolates server failures to prevent one server from affecting others.

Then the sentence to read twice before installing anything: remember that MCP servers can access external services and execute code on your behalf. The same caution applies to extensions, where the marketplace badge means less than people assume — verification confirms publisher identity, which is not a review of the code.

What It Costs, In Mechanisms Rather Than Numbers

Prices and model names change monthly; the mechanics do not. Included usage is consumed per token, not per message, so your model selection affects how quickly your included usage is consumed. It does not accumulate: usage resets monthly with your billing cycle. Unused usage does not roll over. A spend limit is a hard stop rather than a throttle — when a limit is reached, AI features stop for that user until the next billing cycle.

Bringing your own API key changes less than people expect. Custom API keys only work with chat models. Tab completion continues using Cursor's built-in models, the traffic still goes through Cursor's infrastructure, and if your team depends on zero data retention the guidance is explicit: if your team relies on Zero Data Retention, use Cursor's built-in models instead.

When It Breaks, Suspect The Network

Three failures account for most "the AI stopped working" reports, and none of them is the model. Cursor uses HTTP/2 for streaming responses. Some corporate proxies (like Zscaler) block HTTP/2. Over a remote connection the requests do not come from where you think: when using Cursor's Remote SSH extension, AI features run on your local machine and communicate with the remote server for file access. And a disconnected VPN can leave its settings behind, because Cursor can inherit DNS settings from a previously active VPN.

If you do report a bug, know that the privacy setting is captured at submission: Privacy Mode is per-request. Changing your privacy setting does not retroactively affect previous requests.

Where to Go Next

Cursor AI Features picks up where this page stops: what the AI actually does once the setup above is in place — Tab, Inline Edit, Agent and its modes. MCP Fundamentals covers the protocol underneath the servers configured here, rather than Cursor's particular way of loading them. And if you are weighing agentic editors against one another, Claude Code Basics is the same ground in a terminal-first tool.