Cursor appears in Not all AI-assisted programming is vibe coding (but vibe coding rocks) as a popular tool used for Vibe Coding that was not originally designed for it. Simon Willison notes that it was initially intended for professional developers, and treats that origin as the explanation for how little it constrains what generated code can do.
A survey on AI agentic programming, AI Agentic Programming: A Survey of Techniques, Challenges, and Opportunities, describes Cursor as extending IDE-assistant functionality with conversational interaction, memory of previous edits, and structured command execution. Elsewhere in its comparative analysis, the same survey reports Cursor’s underlying model as Claude 3.5 Sonnet or GPT-4, with a 128,000-token default context window and persistent memory implemented as semantic search over project history, and classifies Cursor, in its own taxonomy, as an IDE Assistant that is reactive and tool-using but not multi-turn or adaptive.
Capabilities
The post’s one direct characterisation is comparative: Cursor has far less in the way of safety rails than Claude Artifacts, whose sandbox prevents unreviewed code from reaching the network or from causing harm outside the project.
Karpathy’s account of vibe coding, quoted in the same post, names Cursor Composer driven by a Sonnet model as the setup he was working in, and describes talking to it by voice rather than typing. That is his description of his own workflow at the time rather than a statement about what the product generally offers.
A later post on long-running agents describes Cursor as also shipping background cloud agents: long-running tasks that run on Anysphere’s cloud infrastructure rather than the developer’s own laptop, so an eight-hour refactor or a codebase-wide migration survives a closed lid. A task can be started locally, sent to run in the cloud once it looks like it will take longer, and re-attached to later from another device; each background agent runs in its own isolated git worktree and merges its result back via pull request (see Git worktrees). The same post names Composer 2, described as Cursor’s proprietary frontier coding model, which ships in Cursor 3.
That post also describes the coordination system behind Cursor’s production long-running agents: a Planner/Worker/Judge role split reached after two earlier coordination designs (equal-status agents sharing files with locks, then optimistic concurrency control) each ran into bottlenecks or coordination failures (see Planner-Worker Model). It reports that a GPT model outperformed Opus specifically for extended autonomous work, because Opus tended to stop early and take shortcuts — different models suited to different roles in the same system.
Rule Taxonomy and Evolution in AI IDEs: A Mining and Survey Study treats Cursor as an
AI IDE — an editor built with AI as a core architectural component rather than a plugin
over an existing one — and as one of five such tools that let developers define
AI IDE Rules. Cursor’s rule files live under .cursor/rules/; that study also lists a
legacy .cursorrules format among Cursor’s rule-file locations, noting that such legacy formats were
later superseded by the directory-based mechanism. That study records
Cursor’s release date as 6 April 2023, noting that the date corresponds to the version in which it
transitioned to its VS Code-based architecture.
A 2025 practitioner’s blog post, 谈谈 AI 编程工具的进化与 Vibe Coding, attributes Cursor’s lead over the IDE plugins that preceded it to Context engineering, and gives the codebase index as its first breakthrough. On that post’s account, when a project is opened Cursor splits the codebase into small chunks locally and uploads them to Cursor’s cloud servers, where an embedding model embeds them and they are stored in a cloud vector database; when a question is asked, the prompt is embedded, Turbopuffer runs a nearest-neighbour search, and obfuscated file paths and line ranges are sent back to the client, which reads those code chunks locally. The post says the resulting multi-file context is what lets the model implement cross-file calls, fix bugs spanning several files, refactor whole modules and add features that touch many files. It adds that Cursor lets users @-mention files and folders directly, later added indexing of Git history and of documentation, and supports Rules, which it describes as giving the model general coding rules so that generated code stays consistent with a project’s architecture, style and stack. Separately, it credits Cursor with a proprietary model built for Tab completion, fast and accurate enough that the author reports a high rate of accepting its suggestions.
The same post also relays the other side of the argument over retrieval: that semantic similarity is not the same as the context code actually depends on, and that an index kept in sync through a Merkle tree of file hashes, as Cursor’s is, can return out-of-date code during heavy refactoring or when the index servers are under load. In the author’s experience Cursor and Claude Code perform similarly on small and medium tasks, while Claude Code does far better on large tasks touching more than ten files.
Adoption & Ecosystem
Willison groups Cursor with other popular vibe coding tools, which places a tool built for professionals among the ones newcomers reach for. The safety concern he raises follows from exactly that mismatch: the conditions he sets out for vibe coding — low stakes, care with secrets and private data, hard billing limits — have to be met by the person rather than by the tool when the tool does not enforce them.
In the rule-taxonomy study’s practitioner survey, Cursor was the most widely adopted AI IDE among its 99 respondents, named by 57 — closely followed by the VS Code plus Copilot setup at 51, which that study counts as a plugin arrangement rather than an AI IDE. On the mining side, Cursor also accounted for the largest share of that study’s project dataset: 46 of the 83 projects it analysed, drawn from an initial 18,790 candidates before keyword, rule-file and manual filtering.
The blog post above adds one practitioner’s account of how Cursor is used when the code has to be kept clean: the author reports that Tab completion is the feature they use most in serious engineering work — writing the layering, abstractions, comments and method names by hand and letting completion generate the rest — because in their experience LLMs are poor at abstraction, and that this is often more efficient than repeatedly adjusting a conversation to get the model to output complete code.