What Cursor Is
Cursor is a fork of VS Code that puts AI at the core of the editor rather than bolting it on as an extension. Completion that predicts the line you are starting, inline editing that rewrites a selection from a natural-language instruction, and an agent that carries out changes across many files — AI is present through the whole process of writing code.
The three features are arranged by how much you hand over: from small, fast completion to an agent you delegate a whole task to. Switch tools according to the size of the work.
Install and Setup
- Download and launch: get the Mac / Windows / Linux build from cursor.com. On first launch you can import your VS Code extensions, themes and keybindings.
- Sign in: create an account with Google / GitHub / email. You can start on the free Hobby plan; trial conditions for paid plans change, so check the official page.
- Check models and privacy: decide which model to use and review Privacy Mode, which keeps your code out of training. This matters most for work code.
Three Main Features
Tab completion
As you start typing, Cursor predicts the next edit. Confirm with Tab, reject with Esc. It handles a line or a few lines, and sometimes linked edits in several places.
Cmd+K (inline edit)
Select the code, press Cmd+K, then instruct in natural language: "turn this loop into map," "make the types strict." The fastest way to fix a single block.
Agent
The panel on the right. Ask for work that spans several files ("add an API endpoint and write tests") and the agent plans and executes it, repeating edit, run and check on its own until the task is finished.
How Far the Agent Goes
The agent can run inside your IDE, in the background, or in an isolated cloud environment where you only collect the result. Several subtasks can run in parallel on separate branches.
It is powerful, but agent-generated code is not automatically correct: it can invent functions or use outdated APIs in a plausible-looking way. Keep three habits: read the diff, run the tests, and require approval for dangerous commands.
Models and How to Choose
Cursor is not tied to a single model; you pick from the "model" dropdown in chat or the agent. As of August 2026 the main choices are Anthropic's Claude Opus 5 and Claude Sonnet 5, OpenAI's GPT-5.6 family, Google's Gemini 3.1 Pro and Gemini 3.7 Flash, xAI's Grok 4.6, and Cursor's own Composer 2.5.The lineup changes often, so check the official model list for the current set.
- Anthropic: Claude Opus 5 / Claude Sonnet 5 — strong at code comprehension and long-running work
- OpenAI: the GPT-5.6 family — general-purpose generation and reasoning
- Google: Gemini 3.1 Pro / Gemini 3.7 Flash — good with long context
- xAI: Grok 4.6
- Cursor's own model: Composer 2.5 — tuned for long coding sessions, easy on speed and cost
When in doubt, leave it on Auto: it picks a fitting model for the task and is easier on your usage allowance. Once you are used to it, switch manually — frontier models for hard design work, fast low-cost models for light edits.
Cursor Rules
Write down rules specific to your team or project (naming conventions, the test framework, prohibitions) and the AI takes them into account every time.
| .cursor/rules/*.mdc (Project Rules) | AGENTS.md |
|---|---|
| Structured rules with apply conditions (always / glob / description-based / manual) | A single plain markdown file at the project root; readable with no configuration |
| Good when you want rules split by role | Good for small to mid-size projects that just need something concise |
The single-file .cursorrules is still read for now but is heading toward deprecation; new projects should use Project Rules or AGENTS.md.
Example rules
Strict types (no any) / functions single-responsibility / camelCase naming / write tests with Vitest / always confirm before running destructive commands.
Rule of thumb: if you have explained the same thing three times, it should be a rule.
How Pricing Works
Billing is a credit allowance that draws down as you use models. Each paid plan comes with a monthly credit allowance equal to the plan price, and consumption depends on which model you choose and how heavy the work is.
- Hobby (free): limited completion and agent use — enough to try it
- Pro (USD 20/mo): the standard for individuals, with a USD 20 credit allowance. Auto mode keeps working after the allowance is used up, at no extra charge
- Pro+ (USD 60/mo): a larger allowance than Pro
- Ultra (USD 200/mo): a large allowance for heavy users
- Teams (USD 40/user/mo): central management, SSO, org administration. Enterprise is quoted individually
So using expensive frontier models manually burns the allowance faster, while leaning on Auto keeps costs down. Exact prices, allowances and trial conditions change, so check the official pricing page before you subscribe.
Mastery Tips
- Don't over-trust Tab: predictions can be wrong. Read before you accept, especially types and boundary conditions.
- Cmd+K for local fixes, the agent for big ones: match the tool to the size of the change.
- Ask the agent to plan first: "make a plan and wait for my OK" prevents sprawling changes.
- Pass context with @: @file and @folder to scope the reference, @web to include web information.
Sticking Points
- Credits go fast: big agent tasks on expensive models drain the allowance. Use Auto by default and frontier models when it matters.
- Taking generated code on trust: verify with diff review and tests.
- Confidential code: turn on Privacy Mode; for organizations, combine it with Teams/Enterprise administration.
- Lag behind upstream VS Code: as a fork, it can trail the newest upstream features.
Using with Other Tools
- Avoid double completion: if you use Cursor's Tab, turn other completion tools such as GitHub Copilot off.
- Split by role: Cursor for everyday editing, terminal-driven tools such as Claude Code for large-scale automation.
Summary
Cursor packs a ladder of delegation — completion, inline edit, agent — into a single editor. Billing is usage-based (credit allowance plus Auto), so start small on Hobby or Pro with Auto mode and teach it your conventions with Rules. Next: the complete guides to GitHub Copilot and OpenAI Codex.
Usage tip:Cursor 1.6 lets you drop a .md file into .cursor/commands/ — it shows up as a shareable custom slash command in the Agent input dropdown (Source: https://prompts.ninja/news/cursor-1-6-slash-commands-summarization/)
Usage tip:In Cursor CLI, type `/btw` to ask a side question without derailing the agent's current run (Source: https://x.com/F2aldi/status/2046442265154076704)
Usage tip:In Cursor, you can set Bugbot's review effort level from its dashboard — use high effort for infra/backend changes so it catches more issues (Source: https://x.com/cursor_ai/status/2053892050299597107)
