Depends on two npm CLIs: xcodebuildmcp for build/launch/UI automation and openspec for cross-session spec tracking.
Supports Claude Opus 4.7 for complex tasks and Sonnet 4.6 for faster iteration; model is user-selectable per task.
Build caching kicks in after the first run – scheme, project path, and simulator ID are reused automatically on subsequent tasks.
Pure SwiftUI + Foundation, no external Swift dependencies; API calls use streaming SSE via AnthropicService.swift.
Hacker News Comment Review
The UI automation question is unresolved: it is unclear whether AgentSwift can visually inspect the simulator screen or only verify non-crash behavior, which is a meaningful gap compared to Playwright-style web automation.
The core differentiator over Claude Code + xcodebuildmcp MCP is not yet clearly articulated by the author; the native GUI, built-in build caching, and queued message handling are the most concrete additions visible from the source.
Notable Comments
@trueno: asks directly what AgentSwift adds over Claude Code with xcodebuildmcp plugged in, noting it uses Claude either way.
@jumploops: “without any skills/plugins, it was booting the simulator and running tests to verify something” – draws parallel to Codex behavior where simulator launch does not equal visual validation.
