AgentSwift is an open-source macOS app that runs Claude as an iOS development agent. You describe what you want, and Claude discovers your project structure, edits source files, compiles via xcodebuildmcp, launches a simulator, and runs UI automation to check everything works. The agent tries one automatic fix on failures before handing problems back to you.
The tool supports Claude Opus 4.7 for heavier work and Sonnet 4.6 for faster iteration. Built with pure SwiftUI and Foundation, no external Swift dependencies. Setup requires macOS 26.1 or later, Xcode, Node.js, and an Anthropic API key, plus two npm packages for build automation and spec tracking.
Hacker News commenters immediately raised the obvious question: how different is this from running Claude Code with an Xcode MCP server? The overlap is real. AgentSwift's pitch is a packaged, opinionated workflow in a native macOS app instead of a terminal session. Whether that matters depends on how much you value GUI convenience.
That single auto-fix attempt defines the tool's limits. If the build fails twice, you're on your own. Simple view changes or adding a settings screen? Probably fine. Anything touching Core Data migrations, complex SwiftUI state, or custom build targets and you'll hit that wall fast. The MIT license at least means you can inspect exactly where and why it breaks.