Skills and CLAUDE.md weren't enough to keep Claude's SwiftUI code on track, so we gave it a static-analysis reviewer over MCP

I'm one of the two developers behind this (Nil Coalescing), so take it with the usual grain of salt. The problem: telling Claude how to write good SwiftUI up front doesn't hold. The guidance is loaded at the start, and by the time Claude writes a list or a chart, it has drifted out of focus. More instructions just cost more tokens. What we built: SwiftFairy, a Mac app that runs a local MCP server. Claude sends the code it just wrote, and the app parses it into a partial graph of your project (the changed files plus the declarations they depend on) and checks that graph against a set of rules. It's deterministic static analysis, with no second model. Claude gets back specific lines, what's wrong and how to fix it, and it's much better at acting on "line 42 has this issue" than at remembering a paragraph from earlier in the session. A few things we learned: • A concrete finding tied to a line is followed far more reliably than general guidance. • Returning less text helps. Results are deduplicated and kept short so they don't flood the context. • When the analyzer isn't sure, it says the problem is possible rather than definite, and Claude weighs it accordingly instead of over-correcting. It works with Claude Code, and also with Codex, Cursor and Gemini CLI. It ships with knowledge packs for SwiftUI, Swift Charts and modern Swift. The announcement post (https://nilcoalescing.com/blog/IntroducingSwiftFairy/) has examples. It's free to try for 24 hours with no card required, and current pricing is on the site. It needs macOS 26 on Apple silicon.