Claude Code, Cursor, and similar AI coding agents have gotten genuinely good at writing Swift and SwiftUI. What they usually can't do on their own is the next step: install the app on an iOS Simulator, tap through the actual screens, and check whether it works — because that normally requires a Mac, Xcode, and either a human at the keyboard or a lot of custom scripting to give the agent "eyes and hands" on the running app.
Three real tools now solve this, each for a different situation. If you don't have a Mac at all, a cloud-hosted simulator removes that requirement entirely. If you have a Mac and want the agent testing against something closer to a real device (not just Xcode's simulator), a local virtual iPhone does that. And if you're running several AI agents at once against the same Mac, you need something to hand out simulators without them colliding — which is a narrower problem than it sounds, since a simulator takes real time to boot and can't safely be shared naively.
1. AgentCloud — an iOS Simulator in the cloud, no Mac required
AgentCloud puts the iOS Simulator itself in the cloud and exposes it through MCP (Model Context Protocol), so an agent like Claude Code, Cursor, Windsurf, Gemini CLI, or Codex can install an app onto a real cloud-hosted simulator, see the screen and accessibility tree, tap and swipe through the UI, navigate via deep links, and read back console output — without you owning a Mac at all. Sessions are disposable and secured via OAuth 2.1 with PKCE, with team-scoped access for shared use.
Watch out for: the free tier is genuinely limited (2 Simulator hours plus 30 lifetime build minutes total, not recurring), it's a new service with less track record than an established player like BrowserStack, and it's tightly coupled to the still-young, fast-moving MCP ecosystem.
Pick AgentCloud if: you're developing an iOS app without owning a Mac, or you want your AI agent testing in a disposable cloud environment instead of your own machine.
2. vphone-cli — a real virtual iPhone on your own Mac
Xcode's Simulator doesn't behave exactly like a real device, which matters for certain testing and security work. vphone-cli takes a different approach: using Apple's own Virtualization.framework, it boots an actual virtual iPhone — real iOS, real behavior — directly on your Mac in one command, with support for firmware variants ranging from patchless to fully jailbroken (auto-installing Sileo and TrollStore on jailbroken images). It exposes a control socket for screenshots, touch injection, and gestures, including an MCP integration so an AI agent can drive end-to-end UI tests against that real iOS environment. It's free, open source (MIT), and actively developed with 9.4k+ GitHub stars.
Watch out for: it's macOS-only and requires Apple Silicon plus macOS 15 Sequoia, needs Xcode and the iOS SDK installed, and requires relaxing SIP/AMFI settings — real configuration work, not a one-click install. The jailbreak firmware option is also a gray area depending on what you're using it for.
Pick vphone-cli if: you already have an Apple Silicon Mac, want a genuinely real (not simulated) iOS environment, and are comfortable with the setup — including for security research, which the cloud options above don't really support.
3. Manzanas — a warm pool of simulators for multiple agents at once
Once you're running more than one AI coding agent against the same Mac — say, Claude Code and Codex both testing different branches — a new problem shows up: iOS Simulators are slow to cold-boot, and handing the same one to two agents at once means they trip over each other. Manzanas runs a small daemon that keeps a fleet of simulators warm (via SIGSTOP/SIGCONT rather than cold-booting) and hands them out through a FIFO queue with TTL-bounded exclusive leases, so agents or CI jobs share the Mac without conflicts. The project reports cutting lease-to-live time from about 7 seconds to 0.28 seconds, and adds MJPEG streaming, a browser dashboard, physical iPhone support via WebDriverAgent, and its own MCP integration. It's free and open source (Apache 2.0).
Watch out for: it's a young project without a large-scale production track record yet, it only pays off once you actually have multiple agents or CI jobs contending for simulators, and it needs a dedicated Mac to run the daemon on.
Pick Manzanas if: you're running multiple AI coding agents or CI jobs in parallel against a shared Mac and simulator boot time or contention is actually costing you time.
Which one should you actually use?
| Tool | Where it runs | Best for | Price |
|---|---|---|---|
| AgentCloud | Cloud (no Mac needed) | No Mac available, disposable sessions | Free tier (2h), paid plans |
| vphone-cli | Local (Apple Silicon Mac) | Real iOS behavior, security research | Free, open source |
| Manzanas | Local (Apple Silicon Mac) | Multiple agents/CI sharing one Mac | Free, open source |
These three aren't really competing for the same job. If you don't have a Mac, AgentCloud is the only option here that removes that requirement. If you have a Mac and want your agent testing against something closer to a real device, vphone-cli gets you there for free. And if your actual bottleneck is several agents fighting over one Mac's simulators, Manzanas solves specifically that — you could reasonably run vphone-cli or Xcode's own simulator underneath Manzanas' orchestration layer rather than picking just one of these three.