Cursor

The choice comes down to two profiles: see which one is yours.

Who it is for. Cursor is made for people who write code themselves - every day, by hand, in an editor. It removes the typing routine: completes lines across the project, renames everything at once, and rewrites selected code when asked in words. Agent work is here too: open the agents window, assign tasks across project copies, come back, and accept the results.

No coding at all? Then Cursor is not the first door. It is an editor: it expects you to at least read what it writes. To see how people assign tasks to agents with no experience, visit the vibecoding page - the door is at the end of the section.

Who it is not for. If your code has to move on its own while you are busy with something else, Claude Code and Codex are a better fit: there the agent lives outside the editor.

Just getting started? Haven't written code or built projects yet - start with the page“What is vibecoding”, then come back here for a tool.

If it suits you, install it using section 3. If you're choosing between it and Claude Code, see the decision tree in 7.

Completes as you type. Its signature feature is Tab: while you type, the editor predicts the next lines by looking at the whole project, not a single file. Anysphere has a separate fast model for this - external ones cannot be connected here.

Edits on command. Select a piece of code, say what to do with it, and the editor rewrites it in place. The middle ground between autocomplete and an agent.

Handles tasks with agents. Since version 3.0 (April 2026), agents have a separate window: assign several tasks, each runs in its own copy of the project and does not interfere with the others. There is also /best-of-n: one task goes to several models at once, you compare the results and choose the best.

Handles big projects. Since Sep 10, 2026 Projects is in beta: the coordinator agent does not write code itself, it splits a big task, hands the parts to agents in the cloud and brings back the result for review. The work does not stop when you close your laptop.

Reviews other people's code. Bugbot reads pull requests before a human: since June 2026, an average run takes a minute and a half, finds 10% more bugs than the previous version, and every note has a “fix in Cursor” button.

Takes commands from your phone. Since summer 2026, cloud agents can be launched and accepted from iPhone and iPad: assign a task from your phone, review the diff, merge it.

Stores code itself. On August 17, Cursor launched its own code hosting, Origin: repositories, pull requests, and agents in one place, with two-way GitHub sync. Early beta for paid plans.

In our project. An honest footnote: this site is built by Claude Code and Codex agents, while Cursor is our manual workbench. When the owner digs into code personally, they open Cursor.

Desktop. macOS, Windows, Linux. It is a rebuild of VS Code, so switching is painless: extensions, themes, and keyboard shortcuts import in one click.

Web and phone. Cloud agents run on cursor.com and in the iPhone and iPad app: tasks can be assigned and accepted without a laptop.

Terminal. There is also a CLI for those who live in the console, but it is a side door, not the main one: Cursor's strength is in the editor.

At workstations. JetBrains survey (10,000 developers, January 2026): GitHub Copilot has 29% of workstations, Cursor and Claude Code have 18% each. A year ago, Cursor was half as common in this survey.

As an agent. In September 2026 Artificial Analysis rebuilt its agent index on harder tasks, and the new version has no pairs with Cursor CLI. In the early-July version Cursor CLI held the middle of the table: 62 points with GPT-5.5, 60 with Opus 4.7 and 52 with its own Composer 2.5 Fast. The autonomous Claude Code and Codex ranked higher, and they are a different genre.

Its own model is growing. At launch (March 2026), Composer 2 scored 73.7 on SWE-bench Multilingual, close to the frontier at a price several times lower. The controversy came from elsewhere: the community uncovered that the model was built on the open Chinese Kimi K2.5, and Anysphere acknowledged it in its technical report.

Plans. Hobby is free and needs no card, with small agent limits. Pro costs $20 a month, Pro+ $60, Ultra $200: the plans differ in how much included usage they give. A team seat costs $40, or $120 with five times the agent limits, and annual billing is 20% cheaper everywhere.

Two pools. Cursor's own models, Grok 4.7 and Composer 2.5, spend a separate pool with a noticeably bigger allowance. Claude, Gemini and GPT are billed from the second pool at their API prices, and a heavy model eats it several times faster than a light one. By Cursor's own estimate, daily work with the agent costs $60–100 a month.

For students. The free year of Pro is gone: it closed to new applications on Jun 25, 2026. Discounts are now handed out at student events.

Where the code goes. In Privacy Mode, code is not used for training, neither by Cursor itself nor by model providers; the company has a SOC 2 audit. One nuance: on personal plans, you enable this mode yourself; by default, it is on only for enterprise plans.

What they found. Cursor has had a rich history of vulnerabilities: five remote-code-execution flaws in fall 2025, then DuneSlide in spring 2026, a critical 9.8 out of 10 vulnerability where infected text from a third-party server took over a machine without a single click. Fixed in version 3.0. A separate story was May 2025: fake npm packages named Cursor infected over three thousand machines. The conclusion is simple: update immediately and install only from cursor.com.

The agent sees everything. The agent reads keys, passwords, and .env files as easily as code, and can send them to the model together with context. Keep secrets outside the project; that is user discipline, not a setting.

Money. In summer 2025, the company quietly changed plan limits, triggering a user revolt, a public apology from the CEO, and refunds. Since then, the model has been credit-based and described more honestly; there have been no new scandals.

New owner. Since Aug 14, 2026 Cursor belongs to SpaceX, which has its own model lab xAI with Grok. OpenAI responded to the deal: by Nov 12, 2026 it is winding down the supply of its models to Cursor and will not give it new ones. Claude and Gemini are still in place, and Cursor calls Grok 4.7 its strongest model.