Most AI coding agents try to be jacks of all trades. AgentSwift doesn't. It does one thing: build and test iOS apps without opening Xcode.
The open-source macOS app takes a plain-language description of what you want, then handles everything from editing source files to launching a simulator. Claude does the heavy lifting. The AgentSwift app plugs directly into that pipeline using a command-line tool called xcodebuildmcp for builds and automation, plus openspec from Fission AI for tracking specs across sessions.
If a build fails, the agent tries one fix before handing the error back to you. No infinite loops. Build info gets cached after the first run, so later tasks skip the discovery phase. The whole thing is built with SwiftUI and Foundation with zero external Swift dependencies. You need macOS 26.1+, Xcode, and Node.js. Claude Opus 4.7 for heavy lifting on large codebases. Claude Sonnet 4.6 for quicker iterations on lighter work.
The repo is on GitHub. No word yet on who built it or how it holds up on real production code.
Apple platform development has been underserved by the current wave of AI coding tools. The AgentSwift app is a focused attempt to change that, and the full-cycle approach from code edit to validated simulator build is something general tools still struggle with.