Cursor

Cursor is an AI code editor built around agents that can plan, write, test, and review code across a whole repository. It started as a fork of a familiar desktop editor with unusually good autocomplete, and has grown into something closer to a development platform.

The pitch today is not "finish my line" but "build this feature". Agents run in the editor, on the command line, in the cloud, and from Slack, and they can open pull requests when the work is done.

This review covers the pieces that matter in daily use, the ownership change that reshaped the company in 2026, and the failure modes worth knowing before a team standardises on it.

Cursor Specifications

Now A SpaceX Company: What That Means

On 14 August 2026 SpaceX completed its acquisition of Anysphere, Cursor's maker, in an all-stock deal valued at roughly $60 billion — widely reported as the largest startup acquisition on record. Cursor now operates as a wholly owned subsidiary inside a division branded SpaceXAI.

For day-to-day users the editor did not change overnight. What did change is the strategic context: Cursor is now funded and directed by a company with its own compute and its own model ambitions, and SpaceX models sit alongside OpenAI, Anthropic, and Gemini in the provider list.

If your organisation has procurement opinions about vendor ownership, this is the fact to surface early rather than discover during a security review.

Tab: The Feature People Actually Stay For

Tab is Cursor's inline prediction, and it remains the reason many developers switch. It predicts multi-line edits and, more usefully, the next place you were going to change — following a rename or a signature change through a file rather than completing one line at a time.

It is the least dramatic feature in the product and the one that compounds. Agents get the demos; Tab is what makes the editor feel faster within an hour of installing it.

Because it is inline and instant, it also carries none of the review burden that agent output does. You accept or ignore each suggestion as you type, so a wrong guess costs a keystroke rather than a code review.

The Agent Modes, And When Each Earns Its Keep

Cursor exposes several ways to hand work to a model. Picking the wrong one is the most common reason people conclude that "AI coding does not work".

In-Editor Agent

The Agent reads across your repository, edits multiple files, runs commands, and iterates on failures. This is the right mode for a scoped change you can review in one sitting — adding an endpoint, threading a new field through a stack, fixing a failing suite.

It works best with a clear acceptance test. Given a command that proves success, it will keep going until the command passes.

Composer

Composer handles larger generation work, where you are describing a chunk of functionality rather than editing existing code. It suits greenfield files and scaffolding better than surgical changes in a mature codebase.

Cloud Agents And Automations

Cloud Agents run outside your machine, in parallel, which is what makes fanning several independent tasks out at once practical. Automations trigger agents on a schedule or an event rather than a keystroke.

These are the features that separate Cursor from an editor plugin. They are also the ones that need the firmest review discipline, because work arrives finished rather than as you watch.

Bugbot For Review

Bugbot reviews code and comments on pull requests on GitHub. Treating it as a first reviewer that catches obvious defects — rather than an approver — is the use that holds up.

Steering Cursor With Rules

Cursor Rules are persistent instructions the agent reads on every run: your conventions, the libraries you have standardised on, the patterns you do not want reinvented.

Teams that write real rules get noticeably better output than teams that do not, and the gap widens as a codebase grows. A repository with no rules file gets generic code; one with a good rules file gets code that looks like the surrounding project.

This is the highest-leverage hour of setup available in the product, and it is the step most teams skip.

Security Posture And Enterprise Controls

Cursor lists SOC 2, ISO 27001, ISO 42001, and AIUC-1 certifications. ISO 42001 is the AI management standard specifically, which is still uncommon and matters if your compliance team asks about AI governance rather than general infosec.

For code residency concerns, Cursor offers secure codebase indexing and self-hosted machines. Adoption is broad — the company cites more than half of the Fortune 500, with named users including NVIDIA and Stripe.

Any tool that indexes a private repository is still a data decision. Route it through review before a wide rollout rather than after.

Where Cursor Frustrates Teams

Agent output degrades on very large or unusual codebases. The more your architecture diverges from mainstream patterns, the more the model guesses, and confident wrong code costs more to review than no code at all.

Cost is genuinely hard to predict. Agent-heavy workflows consume far more than autocomplete-heavy ones, and usage varies enormously between developers on the same team, which makes budgeting awkward.

There is also a real skill curve. Handing an agent a vague task produces sprawling diffs; the developers who get the most from Cursor are the ones who scope tightly and verify hard.

Should You Move Your Team To Cursor?

Cursor is the strongest general-purpose AI coding environment available right now, and the gap is widest for teams willing to use agents rather than autocomplete alone. Tab justifies the switch on its own; the agent layer is what changes how work gets planned.

Adopt it if your codebase follows recognisable patterns, your review culture is solid, and someone will own the rules file. Be more careful if your code is unusual, your budget is fixed, or reviews are already a bottleneck — agents produce diffs faster than a stretched team can read them.

Check cursor.com for current plans and limits before committing a team, and confirm the ownership change sits comfortably with your procurement process.