Cursor is a code editor built on top of VS Code, designed around AI agents. It helps write, edit, and understand code: inline autocomplete, project-wide chat, and an agent that can make changes across multiple files at once.
What Is Cursor
Cursor is made by Anysphere. It's a fork of VS Code: extensions, themes, and keyboard shortcuts carry over almost unchanged, so switching from VS Code takes just a couple of minutes.
The key difference from plugins like GitHub Copilot is that the AI is built directly into the environment, not bolted on top. Cursor sees the entire project, works across multiple files in a single request, and can run terminal commands.
Who it's for: developers who want a tight feedback loop with AI right inside the editor; product managers and analysts who build prototypes and tweak code on the fly; and teams that need agentic development with change review built in.
Core Features
Tab — Predicts your next move and suggests multi-line edits, including entire blocks. It factors in broad project context and recent changes.
Agent (Composer) — Handles a task end-to-end: finds the right files, edits several at once, runs commands, and checks its own work.
Chat — Answers questions about the codebase; changes can be applied with a single click via Apply.
Model choice — Cursor's own Composer model plus access to Claude, GPT, and Gemini, depending on the task.
MCP, skills, hooks — Connect external tools, data sources, and project-specific rules.
Parallel and cloud agents — Run multiple agents on isolated copies of the repository without conflicts.
Composer: Cursor's Own Model
📌 Cursor fine-tuned its own agentic model, Composer, on top of Moonshot's open Kimi K2.5 checkpoint, reducing its reliance on Claude, GPT, and Gemini alone.
Version timeline: Composer (October 2025) → 1.5 → 2 → 2.5 (May 18, 2026).
Composer 2.5 plans its work, edits files, runs terminal commands, and verifies its own output. The model adjusts effort to match task complexity — fast answers for simple tasks, longer runs for complex ones.
Why it matters: Composer delivers frontier-level results at a fraction of the cost and with lower latency than comparable models. In the independent Coding Agent Index, version 2.5 ranked third at launch, right alongside Claude Opus 4.7 and GPT-5.5, while costing 10 to 60 times less per task.
Two operating modes:
- Standard — a cost-efficient option for everyday edits
- Fast — quicker, but roughly 6x more expensive for the same token usage
⚖️ Fast mode is enabled by default in Cursor. In practice, many teams use Composer 2.5 Standard as their economical default for routine work, reserving frontier models for the toughest tasks.
Cursor vs. Codex: Two Different Approaches
Cursor and OpenAI's Codex solve similar problems in different ways. Cursor is an interactive environment — it writes code while you watch and steer. Codex is an autonomous agent that runs in the terminal or cloud — you hand it a task and it returns a finished result.
| Dimension | Cursor | OpenAI Codex |
|---|---|---|
| Where it lives | Full IDE (VS Code fork) | Terminal CLI, cloud tasks, VS Code extension; no dedicated IDE |
| Code context | Pre-indexes the repository | Clones the repository into a sandbox for each task |
| Models | Composer, Claude, GPT, Gemini — your choice | OpenAI models by default (GPT-5.x-Codex) |
| Working style | Tight loop with inline Tab suggestions | Delegate the whole task, get a finished result back |
| Token usage | Higher per task | More efficient — claims notably lower token usage per task |
Codex's autonomy levels increase gradually: Suggest (you approve every action) → Auto Edit (edits files automatically, asks before running commands) → Full Auto (reads, writes, and executes inside an isolated sandbox).
The two tools aren't really competing head-to-head. Many developers keep Cursor as their primary environment for interactive work, while reaching for Codex for background or long-running tasks — like migrations and refactors — that don't need constant human input.
Setup and Configuration
- Download Cursor from cursor.com for macOS, Windows, or Linux.
- On first launch, import your VS Code settings — extensions, themes, and keybindings.
- Sign in and pick a model (Composer, Claude, GPT, or Gemini) in the agent panel.
- Set project rules through the Rules section — these act as system instructions for the agent. The current format uses files inside
.cursor/rules; a single.cursorrulesfile is still supported but considered legacy. - Enable codebase indexing so the agent can find the right symbols more accurately and use fewer tokens.
- Connect MCP servers for access to external data and tools.
Example project rules file:
# .cursor/rules/main.md
- Respond in English.
- Don't touch production configs without an explicit instruction.
- Run the linter and tests after making changes.
Pricing and Usage Limits
Since June 2025, Cursor bills based on usage rather than the number of requests: each plan comes with a credit pool tied to underlying model API pricing. A light autocomplete costs almost nothing, while a heavy MAX-mode refactor across a large repository can burn through a monthly allowance in a single day.
| Plan | Monthly Price | What's Included |
|---|---|---|
| Hobby | Free | Limited agent requests and Tab |
| Pro | $20 | $20 of model usage, unlimited Tab |
| Pro+ | $60 | Roughly 3x the Pro limit |
| Ultra | $200 | Roughly 20x the Pro limit, priority access to new features |
| Teams | $40 per user | Shared billing, analytics, SSO |
| Teams Premium | $120 per user | 5x the Standard limit, for heavy users |
| Enterprise | Custom | Usage pools, SCIM, priority support |
On team plans, the usage pool splits into two buckets: first-party (Composer and Auto) and third-party models (Claude, GPT, Gemini).
Which plan fits whom: Pro works for solo, everyday use; Pro+ suits daily agent-heavy workflows; Ultra is built for long, heavy tasks; Teams covers shared billing and collaboration.
Pricing and limits reflect July 2026 figures; Cursor updates them regularly, so check cursor.com/pricing before purchasing.
⚠️ MAX mode and long contexts burn through credits fastest. The most cost-efficient combination is Composer Standard paired with Auto model selection.
Practical Use Cases
- Prototypes and landing pages — spin up a working page from a single prompt, then refine it in place.
- Refactoring and migrations — move a project to a new framework or library with the agent doing the heavy lifting.
- Exploring unfamiliar codebases — ask questions about the code and get a map of dependencies.
- Debugging — the agent spins up the environment, reproduces the bug, and proposes a fix.
Common mistakes to avoid:
- Vague tasks with no clear definition of "done"
- Blindly trusting the agent's output without reviewing changes
- Expecting everything to work in one pass without iteration
Cursor represents a genuinely different philosophy from purely autonomous coding agents: it keeps you in the loop while still automating the tedious parts of development. Whether it becomes your primary environment or a complement to tools like Codex will depend on how your team likes to work — but for anyone serious about agentic coding, it's worth trying firsthand.