Cursor
Overview
Cursor is an AI-first code editor developed by Anysphere, a company founded in 2023. It was initially a research project at Harvard before becoming a commercial product. Cursor combines a traditional IDE (based on VS Code) with AI coding capabilities, enabling developers to write code faster with AI assistance.
Key Products
Cursor IDE
- AI-powered code editor built on VS Code
- Features include:
- AI Chat: Natural language interactions for code explanation, debugging, and generation
- Autocomplete: AI-powered code completion
- Composer (v2.0+): Self-developed code model for faster, more accurate generation
- Cursor Rules: Custom instructions for AI behavior
- Multi-agent: 8 parallel agents in v2.0
- Built-in Browser: Chromium embedded for live testing
- Voice Mode: Speech-to-text control
Cursor Rules
Configuration files that define how Cursor’s AI behaves in specific contexts:
- Always Apply: Global rules for entire codebase
- File-specific: Rules targeting specific files/patterns
- Intelligent Apply: Dynamic loading based on context
- Manual Apply: On-demand via @file references
Version History
| Version | Key Features |
|---|---|
| 0.5 | Initial releases, basic AI autocomplete |
| 0.50 | Major updates, cost-saving optimizations |
| 2.0 | Composer model, 8-agent parallel, speed 4x faster |
Comparison with Other Tools
| Tool | Strengths | Weaknesses |
|---|---|---|
| Cursor | GUI-based, beginner-friendly, good for frontend | Slower than Claude Code |
| Claude Code | CLI-based, more flexible, better for complex tasks | No GUI |
| Windsurf | Good for long tasks | Doesn’t support Claude 4 |
| Devin | Good integration | Complex setup, code quality issues |
Related Companies
- Anysphere: Parent company, founded by Michael Truell and others
- Sasha Rush: Research scientist at Cursor, leads on Composer model
Source
Cursor 2.0 Release Cursor Programming Practices
See Also
[2026-07-20] CursorBench：内部前沿模型评估基准
来源：Fable 5 × CursorBench / Kimi K3（2026-07-17）。
- Cursor 因同时支持所有主流前沿模型与自研模型，成为评估各模型实际性能的”异常中立的裁判”。评估负责人 Nate Schmidt 团队因公开基准分数与开发者真实接受度脱节，自建内部基准 CursorBench。
- CursorBench 刻意还原”混乱、定义不明确”的真实提示（堆栈跟踪 + 一个”修复”；故意指向错误模块测试模型是否质疑用户假设）。正确答案只是门槛，真正评估”模型是否理解了被问的问题”。
- 已知结果：Claude Fable 5 最大努力模式 72.9%（创新高）。Schmidt 观察到 Fable 5 做”全局推理”（vs Opus 的”局部推理”），并给出模型选择启发式（A→B 路径清晰则不需 Fable；不知 B 在哪则 Fable ideal）。
- 团队协作新范式：接触共享代码前，让 agent 先读队友最近提交并标记冲突。