Cursor is worth considering if you already work with code and want AI assistance close to your files, changes and tests. Its strongest fit is a developer or learner who can explain a small task, review the resulting diff and decide whether the software still behaves correctly. Start with the free tier; consider paying only after you know which limits interrupt useful work.
This Cursor AI review is based on official documentation checked on October 4, 2026. We did not run Cursor, measure its speed or test generated code for this review. The assessment below separates documented capabilities from our recommendations. No numerical rating is assigned.
What is Cursor today?
Cursor combines a code editor with an AI coding agent. It is built on the VS Code codebase, with an import path for extensions, settings, themes and shortcuts. That makes it familiar territory for many VS Code users, though it does not guarantee identical behavior from every extension. Check your essential debugger and extensions before moving a real project. See Cursor’s migration documentation (cursor.com).
Who should try it, and who should be cautious?
- Developers making regular, bounded changes: the useful question is whether assistance reduces the work between finding relevant code and reviewing a correct patch.
- Learners who can check the answer: explanations can be useful, but ask why a change works and compare it with the code. An answer that sounds confident is not a passing test.
- Teams with strict data rules: confirm organizational approval and the permitted repositories, models and execution environment first.
- People expecting a finished app from a vague request: budget for requirements, debugging and review. We have no evidence here that Cursor removes those responsibilities.
Which features matter in a small project?
Tab provides suggestions while you edit, including multi-line changes and related edits in other files. It is useful to evaluate separately from Agent: accepting a local suggestion is a different activity from delegating a multi-step task. Tab documentation (cursor.com).
Ask is the documented read-only option for understanding a codebase. Plan Mode helps research and outline a change before implementation. Agent can search files, edit across them and use tools such as the terminal. These are capabilities, not a guarantee that it chooses the right files or fix. Ask, Plan Mode and Agent.
How does it find the relevant code?
The current search documentation describes Instant Grep, with an index built and queried on your machine. Cursor says it does not upload code or paths to build that search index or store codebase embeddings for search. However, a file opened from a match can still be included in a model request. Local indexing therefore does not mean offline AI operation. Current search and indexing documentation (cursor.com).
Our recommendation is to name the failing behavior and relevant files explicitly. For example, “The completed task appears in the open-task list” gives the agent something to investigate. “Improve the app” leaves correctness and scope much harder to judge.
Cursor Free vs Pro: what does it cost?
On October 4, 2026, Hobby was free, with no credit card required and limited Agent access. Published monthly individual prices were Pro $20, Pro+ $60 and Ultra $200, before applicable taxes. Annual billing is a separate commitment; annual effective rates were not verified for this review. Check current pricing (cursor.com) before subscribing.
Pro, Pro+ and Ultra separate usage into Cursor Models and Other Models pools. The documentation does not give a universal fixed number of requests you can assume for each task. Treat older “500 requests” comparisons cautiously. Models and pricing (cursor.com).
Check the Spending dashboard before a long run. Extra charges can apply when on-demand usage is enabled, and a subagent’s model selection can consume a different pool from the parent agent. A subscription price alone is not your spending ceiling. Usage limits and billing controls (cursor.com).
When is Pro worth it? We would upgrade only after recording recurring, useful tasks that the free allowance cannot cover. Compare the bill with time spent reviewing and repairing output, not just the amount of code produced. This review contains no measured savings that would justify a universal recommendation.
Privacy and permissions need separate checks
According to Cursor, Privacy Mode excludes customer data from training and applies provider zero-data-retention arrangements, with exceptions for abuse investigations and designated non-ZDR models. Temporary encrypted caching is also described. Requests still pass through Cursor’s backend, even with your own API key. Privacy Mode therefore does not make AI local-only. See the Data Use policy (cursor.com), dated September 3, 2026. We did not audit its implementation.
Task modes such as Ask and Plan are different from permission Run Modes. The current local modes include Auto-review, Allowlist and Run Everything. Cursor explicitly warns that Auto-review is not a hard security boundary. Do not assume every command will wait for you to approve it. Run Modes documentation (cursor.com).
Reviewing a diff is not necessarily a gate before files change. Cursor’s documentation says normal workspace edits save immediately, and auto-reload may execute them before review. Keep a recoverable baseline and inspect generated changes before running them. Agent security.
A .cursorignore file helps exclude files from normal AI features, but does not block access through terminal or MCP tools. Keep secrets and sensitive client data out of a practice project rather than relying on that file alone. Ignore-file limitations (cursor.com).
Cursor, Claude Code or VS Code with AI?
Compare the workflow you want to keep: navigating the project, describing a change, reviewing files and checking behavior. Try the same bounded task in your existing setup before moving a real project. For a terminal-oriented exercise, our Claude Code beginner guide covers another starting point. We have not run a head-to-head comparison and cannot name a winner for speed, reliability or code quality.
For a side-by-side look at prices, usage limits, where each tool runs and privacy, see our Cursor vs Claude Code comparison.
A useful first evaluation you can actually judge
Download the three-bug practice project (ZIP, with instructions and independent checks). This intentionally broken code was written for the exercise; it is not Cursor output and was not tested in Cursor.
Our proposed test uses one original task-list fixture with three deliberately planted bugs. This protocol has not been run in Cursor. Its expected answers are reviewer-defined acceptance criteria, not results produced by the tool.
- Open-task filtering: for tasks a, b, c, d with completion flags false, true, false, true, filterOpen must return a, c in that order and leave the input unchanged.
- Priority sorting: for priorities 3, 1, 2, 1, sortedByPriority must return b, d, c, a in a new array. The original array and objects must remain unchanged, and equal priorities must retain their order.
- Empty-list percentage: completionPercent must return 0 for an empty list, rather than NaN. The same four-task fixture must still return 50.
- Freeze the baseline and independent tests. Keep the answer key and acceptance suite outside the editable project. Start each case from a fresh copy with all three original bugs; fix only the assigned one.
- Explain first in Ask. Request the current behavior and the relevant line without changing files. Keep the same model, settings and prompt language across the three cases.
- Request one narrow patch. In Agent, allow only the target function and its new regression test to change. Exclude dependencies, network access and unrelated refactoring.
- Review before running generated code. Inspect every changed or added file, preserve the source diff, and check that neither the answer key nor the independent tests changed.
- Run the independent suite. Save baseline and post-edit output, errors and exit codes. In each fresh case, the targeted failure should disappear while the two untouched intentional failures remain. A green agent-written test alone does not establish success.
- Keep the full record. Save before/after code, the entire chat, exact version and model, task and permission modes, displayed usage, retries and human corrections. Report the first attempt separately from any assisted correction.
The planned run uses an existing authorized desktop-editor installation and allowance, with no Cloud Agent, private repository or additional spending. Stop if access, allowance or permissions cannot be verified. A CLI-only test would not establish the desktop editing experience.
Our verdict
Try Cursor for controlled, reviewable coding work. Its documented combination of editing, code search and agents makes it a credible option for that workflow. The reasons to hesitate are equally practical: variable usage, review effort, extension compatibility and the data access each mode requires. The free tier is a sensible starting point; a paid plan should follow evidence from your own tasks.
Scope and currency: official sources were rechecked on October 4, 2026. No installed Cursor version was verified. Generated-code quality, reliability, prompts in languages other than English, migration and performance remain untested. The three-case protocol is preparation for a future test.
