Anthropic added a live iOS Simulator view to Claude Code Desktop, letting its coding agent watch the app it just built run on a virtual iPhone instead of guessing whether the code compiled. The feature, published in Anthropic’s developer documentation, closes a gap that has dogged coding agents since they moved past web apps: verifying a mobile screen actually works. Until now, an agent could write the Swift, but confirming the screen rendered correctly required the developer to open Xcode and look.
The mechanism is specific. When Claude builds, installs, or launches an app inside a simulator, a pane appears on its own, streaming the device screen at a configurable frame rate and resolution. Anthropic says the pane controls the device natively rather than through computer use, the screen and mouse control mode Claude also supports, so the agent’s testing session does not hijack the developer’s monitor or minimize other windows. The CLI version of Claude Code still reaches the simulator through computer use, meaning it takes over the visible screen the way a person clicking a mouse would. The desktop pane is the one built to run alongside a developer’s other work.
Developers keep a hand on the wheel while the agent works. Tapping, swiping, pressing the hardware Home or Lock buttons, and rotating the device all remain available from the pane, and a visible badge marks whenever Claude, rather than the human, is currently driving the screen. A session can run up to four simulated devices at once, each bound to whichever session started it, so parallel Claude sessions building different features do not collide on the same virtual phone.
Access requires consent, granted once per device rather than once per session, and Anthropic discloses that screenshots Claude takes of the simulator are sent to its servers and retained under the account’s normal conversation settings. That is a reason to keep test accounts, not production logins, on any device an agent touches. Two specific actions, opening a URL on the device and building the app through Xcode’s build scripts, still run under the session’s ordinary permission mode rather than the one-time device consent, since both can move data or execute arbitrary scripts on the developer’s Mac.
The setup requirements narrow the audience by design. The feature needs Claude Desktop version 1.24012.0 or later, a Mac, and a working Xcode installation with the iOS Simulator runtime, and it functions only in local sessions; cloud and SSH sessions run on machines that cannot reach a simulator on the developer’s own hardware. That last restriction is structural, not a rollout choice: Apple’s iOS Simulator only runs on macOS, so there is no path to a cross platform version of this pane.
The real significance is the loop it closes. A model that writes correct Swift is a code generator. A model that can then launch the result, watch it render, and decide whether the screen matches the request is closer to an agent that verifies its own output, the same claim Anthropic and its competitors have made for terminal and browser environments for two years. Mobile has lagged because simulators are inherently graphical, with no log output that tells an agent a layout is broken. Wiring a screen stream directly into the agent’s context, rather than routing it through generic screen control, is what makes that verification cheap enough to run on every change.
Teams building iOS features with Claude Code should treat the simulator pane as the new baseline for review before merging agent-authored UI changes, and should audit what test accounts sit on any simulator an agent has permission to drive.
Documented by Anthropic on July 21, 2026, in the Claude Code Desktop documentation for testing iOS apps in the simulator.
