Skin deep. You will clone the interface and then meet the wall. Realistically a long weekend.

What stops a copy of Cursor working?

Retrieval is the entire product. Deciding which twelve files out of four thousand belong in the context window is the difference between a useful edit and a confident wrong one, and there is no library that does it for you. Latency budgets. Autocomplete has to answer in under a few hundred milliseconds or people turn it off permanently, which means speculative execution, aggressive caching and a much smaller model than you wanted. Applying a diff reliably. Getting a model to describe a change is easy; landing it in the right place in a file that has moved underneath you is where copies visibly fail. Cost per user. Every keystroke is potentially an inference call, and the unit economics only work with routing, caching and a lot of tuning.

Can AI copy

Cursor

A code editor with a model wired into the middle of it.

Skin deep

Verdict: Skin deep
Time to copy: a long weekend
Category: ⌘ Dev tools
Source: Proprietary

The prompt

Written to be pasted straight into an agent with no editing. It describes the whole job, including the parts that will go wrong.

Build an AI coding assistant as an editor extension. Chat panel with the current selection and open files as context, plus an inline edit mode that streams a diff and offers accept and reject per hunk. Build a retrieval layer first: index the repository into chunks with embeddings, and at query time combine semantic search with recent-file and import-graph signals to select context under a fixed token budget. Apply edits by matching on surrounding context rather than line numbers, so a stale diff fails loudly instead of corrupting the file. Add a fast autocomplete path with a smaller model, a prefix cache and a hard latency ceiling that abandons the request rather than arriving late.

Open inClaude ↗Terminal agents take it from the clipboard.

What copies cleanly

The shell. VS Code is open source, the extension API is documented, and a chat panel talking to a model API is an evening.
Inline diff previews, accept and reject controls, a file-mention picker.
Something that autocompletes convincingly in a demo video.

Where it stops

Retrieval is the entire product. Deciding which twelve files out of four thousand belong in the context window is the difference between a useful edit and a confident wrong one, and there is no library that does it for you.
Latency budgets. Autocomplete has to answer in under a few hundred milliseconds or people turn it off permanently, which means speculative execution, aggressive caching and a much smaller model than you wanted.
Applying a diff reliably. Getting a model to describe a change is easy; landing it in the right place in a file that has moved underneath you is where copies visibly fail.
Cost per user. Every keystroke is potentially an inference call, and the unit economics only work with routing, caching and a lot of tuning.

What actually protects them

Accumulated retrieval and prompting tuning that does not appear anywhere in the interface.
Inference economics available at their volume and not at yours.
Being where the developer already is, which the editor fork buys and nothing else does.

Editor's note

An agent will hand you a working chat-in-editor extension in a session, and it will feel like you have most of it. You have the visible part. The gap between that and something you would leave switched on for a month is almost entirely retrieval quality, which is invisible, unglamorous, and the whole job.

Somebody already did it

Go and look at the original

We would rather you formed your own opinion than took ours. They built the thing; we only measured it.

cursor.com ↗