Cursor for iOS

Cursor for iOS is a deliberately narrow product, and that narrowness is the point. The App Store listing frames it as a way to 'build, review, and ship software from anywhere,' but what it actually ships is an agent supervisor: you launch work, watch it happen, and approve the result. Nothing in the listing suggests a local editor, and the seed's limitations say outright that you cannot write or edit code directly. Judge it against that job and it looks coherent; judge it as a mobile IDE and it isn't one. The verification story is where Anysphere spent its effort. Agents run in their own virtual machines with full development environments, so they can install dependencies, run tests, and iterate without touching your machine. They emit videos, screenshots, and logs as artifacts — meaning the reviewer's job on a phone becomes watching a two-minute demo instead of reading a diff cold. For a small screen, that is a genuinely better format than a wall of unified diff. Point-and-draw and voice mode are the two interactions that justify the app's existence rather than just shrink the desktop UI. Circling a rendering bug in a screenshot and typing three words gives an agent precise spatial context that text alone struggles to convey. Voice is less differentiated — it's dictation with intent parsing — but on a phone, capturing a feature idea while walking is a real workflow, not a gimmick. Model choice is genuinely open. The listing names access to any frontier model from OpenAI, Anthropic, and Google, plus Cursor's own agentic Composer model tuned for speed and token efficiency. That means you can route a cheap refactor to Composer and a gnarly architectural change to a frontier model without leaving the app. Where it gets thin: this is Cursor-ecosystem software. The seed's own 'not for' list concedes the value drops sharply for developers not already on Cursor desktop. The App Store review thread is the most useful honesty signal available — a reviewer's complaint that local workers opened PRs without asking for confirmation, and that they wouldn't use the developer's SSH keys, drew a direct developer response, and the reviewer later reported the SSH-key issue fixed and re-rated five stars. That's a healthy trajectory, but it also means oversight defaults are a live area you should test in your own repo before trusting it unattended. Beyond GitHub, no integration list is documented in the material reviewed, so treat the app as a GitHub-plus-Cursor-desktop tool for now.