Claude Code Now Drives Your iOS Simulator. This Changes Everything.
The loop I've been in for two years:
AI writes the code. I open the simulator. I tap through the screens. I find three things that broke. I paste the error back. AI fixes it. I tap through again.
Rinse. Repeat. An hour gone before you've actually built anything.
That loop just changed.
What Anthropic actually shipped
On July 21, Anthropic released iOS Simulator integration for Claude Code Desktop — public beta. The pitch: tell Claude to fix the bug and verify it in the simulator. It builds the app, launches it in a simulator pane right next to your chat, taps through screens, reads the UI, finds issues, and iterates — while you watch.
Four devices per session, so you can run iPhone SE and iPad layout checks side by side.
What actually impressed me as a dev
It drives the simulator directly. No macOS Accessibility permissions. No Screen Recording permissions. No screen hijacking. This is the part that matters — previous approaches to this kind of UI automation required you to grant tools access that would be a security audit nightmare on any serious dev machine.
It reads the UI. It doesn't just run a script and hope — it observes what actually rendered and uses that to decide what to do next.
And yes — this works for Flutter apps too. The iOS Runner builds for the simulator like any native app.
The honest caveats
⚠️ macOS + Xcode only, local sessions only. If you're on Linux or Windows, this doesn't help you yet. If your workflow is remote pair programming, you need to rethink how you use it.
⚠️ Simulator ≠ real device. Camera, push notifications, performance profiling — still manual. This is a UI logic and flow tool, not a replacement for device testing.
⚠️ Screenshots go to Anthropic. Never sign into real accounts on a device Claude drives. Test accounts and staging only. This matters significantly if you work on anything regulated — banking, healthcare, anything with PII.
⚠️ Visual iteration burns tokens fast. Every screenshot is context. Every read of the UI is more context. If you're not careful, you've spent $20 in tokens watching Claude tap around a login screen.
The bigger picture
Anthropic has been the most conservative of the frontier labs about shipping dev-facing features. This is not conservative. Driving a live device, reading UI, iterating autonomously — this is the kind of thing that normally gets held back for safety review.
The fact that it's in public beta tells you something about how they see the risk profile: the worst case is you waste tokens on a simulated login loop. That's not nothing. But it's not the kind of risk that requires a 6-month safety evaluation.
Android emulator support is reportedly next. If this works as described, it closes the gap between "AI writes code fast" and "I actually shipped something."
Official docs: lnkd.in/ddngsWh7