Codex CLI for Swift and SwiftUI Development: Xcode MCP Servers, Agent Skills, and Build-Test-Debug Workflows

Swift and SwiftUI development sits at an intersection that makes agentic coding both powerful and tricky. The toolchain is GUI-heavy, simulators consume resources, and the compiler’s type system catches errors that text-based AI tools miss entirely. Three MCP servers — XcodeBuildMCP, Apple’s xcrun mcpbridge, and SwiftLens — now give Codex CLI structured access to the full build-test-debug cycle, SwiftUI preview rendering, and compiler-grade semantic analysis. Combined with community agent skills and a well-crafted AGENTS.md, this turns Codex CLI into a serious iOS and macOS development companion.

The MCP Server Landscape for Swift

XcodeBuildMCP (Sentry)

XcodeBuildMCP, acquired by Sentry in early 2026 ¹, ships as a standalone binary with 82 tools spanning builds, tests, simulator control, LLDB debugging, and UI automation ². It operates via the xcodebuild CLI — no running Xcode instance required — making it ideal for headless agentic workflows.

Key tool categories:

Build and test: build_sim, test_sim, build_device — structured error and warning output
Simulator lifecycle: list_sims, boot_sim, install_app, launch_app
LLDB debugging: debug_attach_sim, set_breakpoint, inspect_variable, execute_lldb_command
UI automation: snapshot_ui, tap_element, swipe, type_text — gesture simulation on running simulators
Dynamic tool loading: tools register on demand to minimise context window consumption ²

# ~/.codex/config.toml
[mcp.xcodebuild]
command = "npx"
args = ["-y", "xcodebuildmcp@latest", "mcp"]

[mcp.xcodebuild.env]
XCODEBUILDMCP_SENTRY_DISABLED = "true"

Apple Xcode MCP (`xcrun mcpbridge`)

Xcode 26.3 shipped a built-in MCP server accessible via xcrun mcpbridge ³. It exposes 20 tools that communicate with Xcode’s process through XPC, providing access to internal state unavailable through command-line tools:

File operations: XcodeRead, XcodeWrite, XcodeGrep — direct project file manipulation
Build and diagnostics: trigger builds, retrieve structured compiler diagnostics with fix-it suggestions
SwiftUI previews: render preview screenshots without manually navigating the canvas ⁴
Swift REPL: ExecuteSnippet runs Swift code in a live REPL session
Documentation search: query Apple’s entire documentation corpus from the agent context

[mcp.xcode]
command = "xcrun"
args = ["mcpbridge"]

Critical requirement: Xcode must be running with a project open. The bridge auto-detects the Xcode PID and connects via XPC ⁴. This makes it complementary to XcodeBuildMCP rather than a replacement — use XcodeBuildMCP for the headless build loop, Apple’s MCP for preview rendering and documentation lookups.

SwiftLens

SwiftLens integrates directly with Apple’s SourceKit-LSP to provide 15 semantic analysis tools with compiler-grade accuracy ⁵. Unlike text-based search, SwiftLens sees code exactly as the Swift compiler does:

Single-file analysis (no index required): swift_analyze_file, swift_summarize_file, swift_get_symbols_overview, swift_validate_file, swift_get_file_imports
Cross-file analysis (requires index): swift_find_symbol_references, swift_get_symbol_definition, swift_get_hover_info
Code modification: swift_replace_symbol_body for AI-driven refactoring
Index management: swift_build_index builds the SourceKit-LSP index store for the project

SwiftLens reports 60–90% reduction in context size compared to sending raw file contents, because it extracts only the relevant symbols ⁵.

[mcp.swiftlens]
command = "swiftlens"
args = ["--project-root", "."]

Composing the Three Servers

Each server occupies a distinct layer. Running all three gives the agent a complete toolchain:

graph TD
    A[Codex CLI Agent] --> B[XcodeBuildMCP]
    A --> C[xcrun mcpbridge]
    A --> D[SwiftLens]
    B --> E[xcodebuild CLI]
    C --> F[Xcode.app via XPC]
    D --> G[SourceKit-LSP]
    E --> H[Build / Test / Simulator]
    F --> I[Previews / Docs / REPL]
    G --> J[Semantic Analysis / Refactoring]

Layer	Server	Requires Xcode Running	Tools
Build automation	XcodeBuildMCP	No	82
IDE integration	xcrun mcpbridge	Yes	20
Semantic analysis	SwiftLens	No	15

Agent Skills for Swift and SwiftUI

The agent skills ecosystem for Swift has matured rapidly. Paul Hudson’s open-source SwiftUI agent skill ⁶ addresses the most common mistakes AI tools make when generating SwiftUI code:

Deprecated API avoidance: steers agents away from foregroundColor() in favour of foregroundStyle(), prevents ObservableObject when @Observable suffices
Performance rules: no work in View initialisers or body properties; break up large views
Accessibility: labelled controls, adequate touch targets, scalable fonts
Swift idioms: localizedStandardContains() over manual string comparison, count(where:) over filter().count

Hudson expanded this into a broader Swift Agent Skills collection covering SwiftUI, SwiftData, Swift Concurrency, and Swift Testing ⁶. Install via:

npx skills add https://github.com/twostraws/swiftui-agent-skill --skill swiftui-pro

Additionally, community skills exist for Liquid Glass adoption (iOS 26 design language), SwiftUI performance auditing, and Swift concurrency diagnostics ⁷.

AGENTS.md Template for Swift Projects

A well-structured AGENTS.md eliminates the most common hallucination patterns. Here is a template targeting Swift 6.2 and iOS 18/macOS 15:

# AGENTS.md — Swift/SwiftUI Project

## Language & Toolchain
- Swift 6.2 (current stable, released March 2026)
- Xcode 26.3+ required
- Minimum deployment: iOS 18.0 / macOS 15.0
- Swift Package Manager for dependency management

## Architecture Rules
- SwiftUI-first; AppKit/UIKit only via Representable wrappers
- Use @Observable macro, NOT ObservableObject/StateObject
- Use NavigationStack (iOS) or NavigationSplitView (macOS)
- Structured Concurrency: prefer async/await over Combine
- Use Swift 6.2 InlineArray where fixed-size buffers are needed
- Use Span for safe buffer access instead of UnsafeBufferPointer

## Build Commands
- Build: `xcodebuild -scheme MyApp -destination 'platform=iOS Simulator,name=iPhone 16'`
- Test: `xcodebuild test -scheme MyApp -destination 'platform=iOS Simulator,name=iPhone 16'`
- SwiftPM: `swift build` / `swift test`

## Anti-Hallucination Rules
- Do NOT use APIs deprecated before iOS 18 (List { ForEach } not List(items))
- Do NOT use foregroundColor() — use foregroundStyle()
- Do NOT generate UIKit storyboard code
- Do NOT assume Combine is required for reactive patterns
- Do NOT use AnyView — prefer concrete View types or @ViewBuilder
- Swift 6.2's approachable concurrency: MainActor is default, do NOT add redundant @MainActor annotations

## Testing
- Use Swift Testing framework (#expect, @Test) not XCTest assertions
- ViewInspector for SwiftUI unit tests
- XcodeBuildMCP screenshot tool for visual verification

Workflow Patterns

Pattern 1: SwiftUI Screen Generation with Preview Verification

This workflow uses XcodeBuildMCP for the build loop and Apple’s MCP for preview screenshots:

codex --model o3 \
  "Create a settings screen with NavigationStack,
   grouped List sections for Account, Notifications, and Privacy.
   Build for iOS Simulator, capture a preview screenshot,
   verify layout matches the specification."

The agent writes the SwiftUI view, triggers build_sim via XcodeBuildMCP, then uses xcrun mcpbridge to render a preview screenshot — all without leaving the agentic loop ⁴.

Pattern 2: Codebase-Wide Refactoring with SwiftLens

SwiftLens excels at large-scale refactoring where text-based find-and-replace breaks:

codex --model o3 \
  "Find all types conforming to ObservableObject using SwiftLens.
   Migrate each to @Observable macro.
   Update all call sites that use @StateObject to @State.
   Validate with swift_validate_file after each change."

SwiftLens’s swift_find_symbol_references traces every usage with compiler accuracy, and swift_replace_symbol_body performs surgical edits ⁵.

Pattern 3: Batch Test Generation

codex exec --model o4-mini \
  --input-file views.txt \
  "For the SwiftUI view at {line}, generate a Swift Testing
   @Test function covering the primary user interaction.
   Use ViewInspector for view hierarchy assertions."

Pattern 4: Dependency Audit with Swift Package Manager

codex --model o4-mini \
  "Run 'swift package show-dependencies --format json'.
   For each dependency, check the latest version on Swift Package Index.
   Flag any dependency more than one major version behind.
   Output a markdown report."

Model Selection for Swift Tasks

Task	Recommended Model	Rationale
Architecture design, complex concurrency	o3	Reasoning through actor isolation, Sendable conformance
SwiftUI view generation	o4-mini or GPT-5.5	Sufficient for template-driven UI code ⁸
Test generation, boilerplate	o4-mini	Cost-effective for repetitive patterns
Performance auditing	o3	Requires understanding of SwiftUI rendering pipeline
Xcode project configuration	o4-mini	Scheme and target manipulation is formulaic

GPT-5.5 is now available in Codex and handles most SwiftUI work well, with improved diagnostics interpretation reducing the need for specialised concurrency skills ⁸.

Sandbox Considerations

Codex CLI’s sandbox creates specific challenges for Apple platform development:

Simulator access: XcodeBuildMCP launches simulators via xcodebuild, which requires the full Xcode installation in the sandbox environment. The simctl commands work headlessly but UI automation requires a display server ²
Derived Data: warm the DerivedData cache before starting agentic sessions. Cold builds for non-trivial iOS projects can exceed Codex CLI’s default timeout
Code signing: disable signing for simulator-only development (CODE_SIGN_IDENTITY="" CODE_SIGNING_REQUIRED=NO) to avoid keychain access issues in sandboxed environments
SwiftPM cache: pre-resolve dependencies with swift package resolve to avoid network fetches during sandboxed execution
Preview rendering: xcrun mcpbridge requires a running Xcode instance, which means preview-based verification workflows need a graphical environment — not available in pure headless sandboxes

Limitations

Training data lag: Swift 6.2 shipped in March 2026 ⁹. Models may not fully understand InlineArray, Span, or the approachable concurrency defaults. The AGENTS.md anti-hallucination rules mitigate this
Compose preview requires Xcode: unlike Android’s CLI-only preview rendering, SwiftUI previews require a running Xcode instance via xcrun mcpbridge. XcodeBuildMCP’s snapshot_ui captures running simulator screenshots as an alternative ²
SwiftLens requires macOS: SourceKit-LSP integration means SwiftLens cannot run on Linux CI agents ⁵
XcodeBuildMCP tool count: 82 tools can consume significant context. Use dynamic tool loading to register only the tools needed for the current task ²
Agent skill conflicts: multiple skills advising on the same API (e.g., competing SwiftUI patterns) can produce contradictory guidance. Pin a single authoritative skill per domain
Xcode single-instance: xcrun mcpbridge connects to one Xcode instance. Multi-project workflows need scheme switching, not multiple Xcode windows

Citations

Sentry Blog, “Sentry acquires XcodeBuildMCP”, 2026. https://blog.sentry.io/sentry-acquires-xcodebuildmcp/ ↩
XcodeBuildMCP GitHub (getsentry), v2.3.2 documentation. https://github.com/getsentry/XcodeBuildMCP ↩ ↩² ↩³ ↩⁴ ↩⁵
Jason Moon, “Xcode 26.3 Makes Your IDE an MCP Server — Any AI Agent Can Now Build iOS Apps”, March 2026. https://jasonmoon.dev/blog/2026-03-05-xcode-26-mcp-agentic-coding ↩
Blake Crosley, “Two MCP Servers Made Claude Code an iOS Build System”, 2026. https://blakecrosley.com/blog/xcode-mcp-claude-code ↩ ↩² ↩³
SwiftLens GitHub, “Deep Semantic Swift Code Analysis for AI Models”. https://github.com/swiftlens/swiftlens ↩ ↩² ↩³ ↩⁴
Paul Hudson, “SwiftUI Agent Skill — Write better code with Claude, Codex, and other AI tools”, Hacking with Swift. https://www.hackingwithswift.com/articles/282/swiftui-agent-skill-claude-codex-ai ↩ ↩²
OpenAI Developers, “Build for iOS — Codex use cases”. https://developers.openai.com/codex/use-cases/native-ios-apps ↩
OpenAI Developers, “Codex Changelog — May 2026”. https://developers.openai.com/codex/changelog ↩ ↩²
Swift.org, “Swift 6.2 Released”, March 2026. https://www.swift.org/blog/swift-6.2-released/ ↩

Codex CLI for Swift and SwiftUI Development: Xcode MCP Servers, Agent Skills, and Build-Test-Debug Workflows

The MCP Server Landscape for Swift

XcodeBuildMCP (Sentry)

Apple Xcode MCP (xcrun mcpbridge)

SwiftLens

Composing the Three Servers

Agent Skills for Swift and SwiftUI

AGENTS.md Template for Swift Projects

Workflow Patterns

Pattern 1: SwiftUI Screen Generation with Preview Verification

Pattern 2: Codebase-Wide Refactoring with SwiftLens

Pattern 3: Batch Test Generation

Pattern 4: Dependency Audit with Swift Package Manager

Model Selection for Swift Tasks

Sandbox Considerations

Limitations

Citations

Apple Xcode MCP (`xcrun mcpbridge`)