Cursor

What is Cursor in 2026?

Cursor is an AI-native code editor forked from VS Code, built by Anysphere, and the most-used commercial AI IDE for professional developers. The 3.0 release on 2 April 2026 reframed the product around the Agents Window — many agents in parallel across local repos, git worktrees, remote SSH hosts, and Cursor's own cloud VMs. The default model is now Composer 2, Anysphere's in-house coding model that landed on 19 March 2026 and prices below frontier models while scoring 61.3 on CursorBench (up from 44.2 for Composer 1.5) and 73.7 on SWE-bench Multilingual.

What's new in Cursor 3?

Cursor 3.0 introduced three things worth knowing about. The Agents Window (⌘+Shift+P → Agents Window) runs multiple agents at once across environments — one in your local checkout, one in a worktree, one on an SSH dev box, one in a cloud VM — with Agent Tabs that let you watch them side-by-side or in a grid. Two new commands ship in the same release: /worktree spins up an isolated git worktree per task so agents stop stepping on each other, and /best-of-n runs the same task across multiple models in parallel and lets you pick the winner. Design Mode (⌘+Shift+D) lets you click a UI element in the in-editor browser and add it directly to chat with ⌘+L. The follow-up 3.2 release on 24 April 2026 added /multitask for async subagents and improved multi-root workspace handling.

What is Composer 2?

Composer 2 is Cursor's in-house coding model, trained with continued pretraining followed by reinforcement learning on long-horizon coding tasks. It is now the default model in the editor. The pitch is the cost-quality ratio: Composer 2 ships at $0.50 / $2.50 per million input/output tokens (standard) or $1.50 / $7.50 (Fast, which trades cost for latency, not intelligence) — roughly an order of magnitude cheaper than Claude Sonnet 4.6 ($3 / $15) and still 61.7 on Terminal-Bench 2.0 against Composer 1.5's 47.9. Anysphere's framing is that for the bulk of edit-test-fix loops you do not need a frontier model on every turn.

Composer 2 vs routing to Claude / GPT-5

The honest split: Composer 2 wins on routine in-loop work (tab completions, single-file refactors, test generation, lint-driven cleanups) where the cost-per-turn matters and the agent will run dozens of turns. Claude Sonnet 4.6 still wins on multi-hour autonomous sessions where the agent has to stay on-task across thousands of tool calls — Anthropic's harness tuning shows up there. GPT-5 with reasoning.effort: high wins on hard algorithmic puzzles but adds five-to-fifteen seconds of latency per turn, which compounds badly inside an agent loop. The pragmatic default in Cursor today is Composer 2 for the inner loop, escalate to Sonnet 4.6 or Opus 4.7 when you hit something that visibly stumps it. The full lineup with pricing and SWE-bench numbers is in the AI & LLM coding model comparison.

Cloud Agents — what runs while your laptop is closed

Cloud Agents (formerly Background Agents) execute in dedicated cloud VMs that clone your repo, work on a separate branch, and push back for handoff. Per Cursor's docs: "You can run as many agents as you want in parallel, and they do not require your local machine to be connected to the internet." You start them from cursor.com/agents on any device, from the Cloud dropdown in the desktop app, by commenting @cursor on a GitHub PR or issue, by @cursor in Slack, by @cursor in Linear, or via the Cursor API. The VM means each agent can build, run the test suite, and verify its own work before opening a PR rather than emitting code into the void.

A real multi-agent workflow

The Agents Window earns its keep when the work is genuinely parallel. Concrete setup: agent A runs in a worktree and refactors a service module; agent B runs in the main checkout and updates the integration tests against the new shape; agent C runs as a Cloud Agent kicked off from a GitHub issue and works through a backlog of small bug fixes against a separate branch. You sit at the Agents Window, glance at three tabs, and merge the ones that come back green. Where this falls apart: tasks that share state, tasks where you cannot tell at a glance whether the diff is correct, and tasks where one agent's output is another's input. Parallel agents on coupled work is theatre — slower than one agent doing both jobs in sequence, because you spend the time merging instead of reviewing. Three agents on three independent worktrees is real leverage.

How does Cursor compare to Claude Code and Windsurf?

Three different bets on the same problem. Claude Code is terminal-first — no editor chrome, runs anywhere you have a shell, and leans into git and CLI workflows. It does not have an Agents Window equivalent in the editor sense, but its background mode plus Opus 4.7 on the 1M-context tier covers the same long-horizon work. Windsurf bets on Cascade, a single deeply-integrated agent rather than parallel agents — the trade-off is fewer coordination tabs but no parallelism gain when you genuinely have parallel work. Cursor's edge is the Agents Window itself plus Composer 2's pricing; Claude Code's edge is harness depth around the model; Windsurf's edge is the focused single-agent UX. The deeper view is in agentic coding workflows and the Shapiro five autonomy levels.

Key features

• •Agents Window — many agents in parallel across local, worktrees, SSH, and cloud VMs
• •Cloud Agents triggered from Slack, GitHub, Linear, mobile, or the API; run with the laptop closed
• •Composer 2 as the default in-house model; Claude, GPT-5, Gemini still selectable per-request
• •/worktree and /best-of-n slash commands for isolated and competitive runs
• •Design Mode — click a UI element in the in-editor browser and pipe it into chat
• •Tab completions, ⌘+K inline edits, and codebase-aware chat (the original Cursor surface)
• •.cursor/rules/*.mdc for project-specific instructions (the legacy single-file .cursorrules still works)
• •Full VS Code extension compatibility

Pricing

Composer 2 token pricing if you bring your own API path: $0.50 input / $2.50 output per million tokens (standard), $1.50 / $7.50 (Fast). Verify on cursor.com/pricing — model tiers and surcharges change.

Example .cursor/rules

---
description: TypeScript + React conventions
globs: ["**/*.ts", "**/*.tsx"]
alwaysApply: true
---

You are a senior TypeScript developer.
Use React 19 with functional components.
Always use named exports, never default exports.
Use Tailwind CSS for styling.
Write tests with Vitest.
Prefer early returns over nested conditions.

Extensions & customisation

Pointing Cursor at a local model

Cursor accepts an OpenAI-compatible endpoint as a custom provider, so you can route edits and chat to a local Qwen2.5-Coder running on Ollama. The catch is that Cursor's backend cannot reach localhost — you need an ngrok tunnel and the /v1 suffix on the base URL. The Local LLMs for coding primer walks through the full setup, including the model picks for 12 GB / 16 GB / 24 GB cards.

Terminal-native alternative

If your day is mostly multi-file refactors and shell-driven work, pair Cursor with a terminal agent like OpenCode — open-source MIT, BYO-model across 75+ providers, and reuses your existing GitHub Copilot subscription via /connect. Many teams keep Cursor for inline edits and Composer 2, and reach for OpenCode when they want a terminal-first agent that runs anywhere.