
I kept seeing Cursor mentioned online without knowing what it was, so I installed it on my Mac and wrote down what it does. The short version: it is VS Code, forked, with a language model wired into three places where the editor used to be dumb. Everything else about it is the editor you already know, which is most of why switching costs so little. This is the note as of April 2025; the product changes monthly.
Nothing about the editor is new, on purpose
Cursor is built from VS Code’s source. Any VS Code project opens in it unchanged, the extensions and settings import in one click, and the keybindings are the ones your hands already have. That is a deliberate choice: the pitch is not a new editor but the old one with a model in it, so the cost of trying it is an afternoon rather than a month of relearning. It ships for Mac, Windows and Linux.
The model shows up in three places
- Completion. As you type, it proposes the next few lines, not the next token, and the proposals follow the style of the surrounding code. Accepting is one key. This is the feature that runs all day, and the one that decides whether the tool feels like help or noise.
- Inline editing. Select a block, describe the change, and Cursor rewrites it in place, showing a diff you accept or reject hunk by hunk. The same prompt box takes a file or a symbol as context, so a rename or a refactor across a module is one instruction instead of a search-and-replace.
- Chat that has read the code. A side panel where you ask for an explanation, a fix or a plan, with the open file, a chosen file, or the indexed codebase as context. The difference from pasting into a chat website is that the model sees the project rather than the snippet, and its answer can reference the actual function two directories over.
What it changes, and what it does not
The three features are the same model applied at three scopes: a line, a block, a project. What changed for me is where the typing goes. Boilerplate, test scaffolds and the third variant of a function I have already written twice come from the completion, and the time goes into reading what it produced. What does not change is the reading: a suggestion that compiles is not a suggestion that is right, and the editor gives no signal for the difference. The model is fastest exactly where I am least careful, which is the trade to be aware of before turning it on.
Familiar. Editor. Smarter. Autocomplete. Chat. Knows. Your. Codebase. Switching. Costs. Little.
References
- Cursor and its documentation
- AI-assisted blog posts on this blog, the same trade for prose