Native macOS app that runs an autonomous Claude-powered agent to discover, implement, build, and validate Xcode projects without touching Xcode.
Key Takeaways
Multi-phase agentic loop: Discover, Implement, Build, Launch/Validate, Archive – driven by Claude Opus 4.7 or Sonnet 4.6 depending on task complexity.
Depends on xcodebuildmcp (build, launch, UI automation) and openspec (cross-session spec tracking) as npm globals; no external Swift dependencies.
Architecture is minimal: five Swift files covering the UI, Anthropic streaming SSE client, and a ToolExecutor running bash, read_file, and write_file.
Scheme, project path, and simulator ID are cached after the first build run; subsequent agent calls skip discovery and go straight to implementation.
Error escalation is capped at one auto-fix attempt; the agent surfaces the failure to the user rather than looping.
Hacker News Comment Review
The central open question from commenters: what does AgentSwift add over Claude Code with an xcodebuildmcp server plugged in directly, given it already uses Claude under the hood.
No substantive counter-argument or author response yet, so the differentiation case – native macOS UI, message queuing, build caching, guided simulator selection – has not been defended publicly.
Notable Comments
@trueno: asks directly whether AgentSwift offers “additional guardrailing or features” vs. Claude Code plus xcodebuildmcp – the core positioning question left unanswered.