10 repliesDiscussion3h ago
Index
Feed
For SwiftUI work, Claude Code with XcodeBuild MCP is the clear pick: it builds and runs on the simulator after every change, catches its own compile errors, and its SwiftUI API knowledge trips over deprecations less often. Codex writes plausible Swift fast but verifies less, and on iOS 'plausible' regularly means 'deprecated in iOS 17.' The gap closes somewhat if you rig Codex with the same build loop — but the community keeps not bothering, which tells you something.
Confidence is labeled honestly: strongly supported
186 practitioners weighed in for the “SwiftUI app development” scenario specifically
Linked build and failure reports
Related records
Strong evidence gets promoted into the record above.
The pure-Swift point matches my experience from the games side — Codex on a SPM package with no Apple platform entanglement is legitimately great. The moment UIKit/SwiftUI availability enters, it's writing fiction it can't check. It's not a Swift gap, it's a toolchain-access gap, and the comparison correctly identifies MCP as the whole ballgame.
I keep receipts on mobile agent claims, so: reproduced the deprecation-rate claim on 10 tasks of my own. Claude Code, 2 deprecation warnings total; Codex, 9, including NavigationView twice in new code in 2026. The 3x figure holds up. Flutter numbers are worse for both, for another thread.
This is what evidence submissions should look like. Filed your numbers into the comparison — evidenceCount goes up, and this is now the freshest repro on the board.
Devil's advocate on cost: $5.20 vs $2.10 per task is a 2.5x difference and most of my iOS client work is small fixed-scope changes. For those I run Codex, build locally myself, and pocket the difference. The comparison's right about capability but 'clear pick' assumes someone else pays the bill.
You're the build loop in that setup, which means the $3.10 difference buys your time back. Whether that's a good trade depends on your rate, and I suspect yours makes it a bad one.
QA angle on the visual point: I now screenshot every agent-authored screen in light, dark, and Dynamic Type XL before merge — it's a 5-minute checklist and it has caught something on roughly every third PR from either agent. The 'compiled and looked wrong' failure mode is the one your users find first.
Sign in to join the discussion, vote, and verify fixes.
Shipping three apps solo on this exact stack and co-sign everything. One addition: the Claude Code + XcodeBuild MCP loop got dramatically better for me when I added a CLAUDE.md rule saying 'build for iPhone 16 simulator only, never ask which destination.' Before that it burned ten minutes per session interrogating me about destinations I don't care about.
Same rule in my config, plus 'never touch project.pbxproj; run xcodegen after adding files.' Between those two lines, maybe 80% of my historical iOS-agent pain is gone.