Production-grade MCP server for Xcode workflows — optimized for AI agents with accessibility-first iOS automation
XC-MCP makes Xcode and iOS simulator tooling accessible to AI agents through intelligent context engineering. V3.0.0 adds platform-native defer_loading support — Claude's tool search automatically discovers tools on-demand, minimizing baseline context overhead while maintaining full 29-tool functionality.
Traditional Xcode CLI wrappers dump massive output that exceeds MCP protocol limits:
simctl list: 57,000+ tokens (unusable in MCP context)- Build logs: 135,000+ tokens (catastrophic overflow)
- Screenshot-first automation: 170 tokens per screen, 2000ms latency
- No state memory between operations
V3.0.0 Architecture:
Platform-native defer_loading on all 29 tools
├─ Claude's tool search discovers tools automatically
├─ Tools loaded on-demand (minimal baseline overhead)
├─ Accessibility-first workflow (50 tokens, 120ms vs 170 tokens, 2000ms)
└─ Workflow tools for common operations (fresh-install, tap-element)
Token Efficiency Evolution:
| Version | Baseline Tokens | Total Tools | Architecture | Context Available |
|---|---|---|---|---|
| Pre-RTFM (v1.2.1) | ~45k | 51 | Individual tools | 3.9% (155k) |
| V1.3.2 (RTFM) | ~30k | 51 | Individual + RTFM | 1.5% (170k) |
| V2.0.0 | ~18.7k | 28 | Routers + Full Docs | 9.3% (181k) |
| V3.0.0 | ~0 | 29 | Platform defer_loading | 100% (200k) |
Key Improvements (V3.0.0):
- ✅ Platform-native defer_loading - All tools deferred; Claude discovers on-demand
- ✅ Workflow tools - High-level abstractions for common operations
- ✅ Zero baseline overhead - Platform handles tool discovery
- ✅ Accessibility-first automation (3-4x faster, 3-4x cheaper than screenshots)
- ✅ Progressive disclosure (summaries → cache IDs → full details on demand)
- ✅ 60% test coverage with comprehensive error handling
# Install globally
npm install -g xc-mcp
# Or run without installation
npx xc-mcpMCP Configuration (Claude Desktop):
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"xc-mcp": {
"command": "npx",
"args": ["-y", "xc-mcp"]
}
}
}Minimal Mode (for Claude Code and other clients that don't support defer_loading):
{
"mcpServers": {
"xc-mcp": {
"command": "npx",
"args": ["-y", "xc-mcp", "--mini"]
}
}
}The --mini flag reduces tool descriptions from ~18.7k tokens to ~540 tokens (~97% reduction). Use rtfm for full documentation on-demand.
Build-Only Mode (for build-focused workflows without UI automation):
{
"mcpServers": {
"xc-mcp": {
"command": "npx",
"args": ["-y", "xc-mcp", "--build-only"]
}
}
}The --build-only flag loads only 11 tools (vs 30): xcodebuild tools, simctl-list, cache, and system tools. Excludes IDB/UI automation and workflow tools. Combine with --mini for maximum reduction: ["--mini", "--build-only"].
XC-MCP returns concise summaries first, with cache IDs for on-demand detail retrieval:
Example: Simulator List (96% token reduction)
// 1. Get summary (2,000 tokens vs 57,000 raw)
simctl-list({ deviceType: "iPhone" })
// Returns:
{
cacheId: "sim-abc123",
summary: { totalDevices: 47, availableDevices: 31, bootedDevices: 1 },
quickAccess: { bootedDevices: [...], recentlyUsed: [...] }
}
// 2. Get full details only if needed
simctl-get-details({
cacheId: "sim-abc123",
detailType: "available-only",
maxDevices: 10
})Example: Build Operations
// 1. Build returns summary + buildId
xcodebuild-build({ projectPath: "./MyApp.xcworkspace", scheme: "MyApp" })
// Returns:
{
buildId: "build-xyz789",
success: true,
summary: { duration: 7075, errorCount: 0, warningCount: 1 }
}
// 2. Access full logs only when debugging
xcodebuild-get-details({ buildId: "build-xyz789", detailType: "full-log" })Discovery Workflow:
// 1. Browse tool categories
rtfm({ categoryName: "build" })
// Returns: List of build tools with brief descriptions
// 2. Get comprehensive docs for specific tool
rtfm({ toolName: "xcodebuild-build" })
// Returns: Full documentation with parameters, examples, related tools
// 3. Execute with consolidated operations
xcodebuild-build({ scheme: "MyApp", configuration: "Debug" })Why RTFM?
- Tool descriptions: <10 words + "See rtfm for details"
- Full docs retrieved only when needed
- 80% token savings vs traditional verbose MCP servers
Before V2.0: 21 individual tools
simctl-boot, simctl-shutdown, simctl-create, simctl-delete,
simctl-erase, simctl-clone, simctl-rename, simctl-install,
simctl-uninstall, simctl-launch, simctl-terminate...V2.0: 6 consolidated routers
simctl-device({ operation: "boot" | "shutdown" | "create" | "delete" | "erase" | "clone" | "rename" })
simctl-app({ operation: "install" | "uninstall" | "launch" | "terminate" })
idb-app({ operation: "install" | "uninstall" | "launch" | "terminate" })
cache({ operation: "get-stats" | "get-config" | "set-config" | "clear" })
persistence({ operation: "enable" | "disable" | "status" })
idb-targets({ operation: "list" | "describe" | "connect" | "disconnect" })Result: 40% token reduction through shared parameter schemas and unified documentation.
XC-MCP promotes accessibility-first automation because it:
- Encourages better apps: Developers building accessible UIs benefit all users (screen readers, voice control, assistive technologies)
- Enables precise AI interaction: Semantic element discovery via accessibility tree vs visual guesswork from screenshots
- Improves efficiency: 3-4x faster execution, 3-4x cheaper token cost
- Reduces energy usage: Skip computationally expensive image processing entirely
| Approach | Tokens | Latency | Use Case |
|---|---|---|---|
| Accessibility Tree | ~50 | ~120ms | Rich UIs with >3 tappable elements |
| Screenshot Analysis | ~170 | ~2000ms | Minimal UIs with ≤1 tappable element |
| Efficiency Gain | 3.4x cheaper | 16x faster | When accessibility sufficient |
// 1. ALWAYS assess quality first
accessibility-quality-check({ screenContext: "LoginScreen" })
// Returns:
{
quality: "rich" | "moderate" | "minimal",
recommendation: "accessibility-ready" | "consider-screenshot",
elementCounts: { total: 12, tappable: 8, textFields: 2 }
}
// 2. Decision branch based on quality
if (quality === "rich" || quality === "moderate") {
// Use accessibility tree (faster, cheaper)
idb-ui-find-element({ query: "login" })
// Returns: { centerX: 200, centerY: 400, label: "Login" }
idb-ui-tap({ x: 200, y: 400 })
// Precise coordinate-based interaction
} else if (quality === "minimal") {
// Fall back to screenshot (last resort)
screenshot({ size: "half", screenName: "LoginScreen" })
// Visual analysis when accessibility insufficient
}Why This Matters:
- For Users: Encourages inclusive app development benefiting everyone
- For AI Agents: Precise semantic targeting vs visual pattern matching
- For Efficiency: 50 tokens (accessibility) vs 170 tokens (screenshot)
- For Speed: 120ms (accessibility) vs 2000ms (screenshot)
- For Energy: Skip image encoding/decoding/analysis entirely
accessibility-quality-check: Rapid assessment without full tree query
- Returns:
rich(>3 tappable) |moderate(2-3) |minimal(≤1) - Use case: Decision point before screenshot vs accessibility
- Cost: ~30 tokens, ~80ms
idb-ui-find-element: Semantic element search by label/identifier
- Returns: Tap-ready coordinates (centerX, centerY) with frame boundaries
- Use case: Find specific button, field, or cell without visual analysis
- Cost: ~40 tokens, ~120ms
idb-ui-describe: Full accessibility tree with progressive disclosure
- Operation
all: Summary + uiTreeId for full tree retrieval - Operation
point: Element details at specific coordinates - Use case: Discover all interactive elements, validate tap coordinates
- Cost: ~50 tokens for summary, ~500 tokens for full tree
XC-MCP V3.0 adds the defer_loading: true flag to all 29 tool registrations. Claude's platform-native tool search automatically:
- Discovers tools on-demand — No custom tool-search implementation needed
- Loads tools when relevant — Based on conversation context
- Minimizes baseline overhead — Zero tokens at startup
Use rtfm to get comprehensive documentation for any tool:
// 1. Browse tool categories
rtfm({ categoryName: "build" })
// Returns all build-related tools with descriptions
// 2. Get comprehensive docs for specific tool
rtfm({ toolName: "xcodebuild-build" })
// Returns full documentation with parameters, examples, related tools
// 3. Execute with discovered parameters
xcodebuild-build({ scheme: "MyApp", configuration: "Debug" })Default (V3.0.0): All tools have defer_loading enabled
# Platform discovers and loads tools automatically
# Zero baseline token overheadDisable defer_loading (for debugging/testing):
# Set environment variable to load all tools at startup
export XC_MCP_DEFER_LOADING=false
# All 29 tools loaded immediately (~18.7k tokens)
# Useful for: Testing, debugging, MCP client compatibilityXC-MCP provides 2 high-level workflow tools that combine common operations into single steps:
Combines accessibility quality check + element search + tap into one operation:
workflow-tap-element({
elementQuery: "Login",
screenContext: "LoginScreen",
inputText: "user@example.com", // optional: type after tap
verifyResult: true // optional: screenshot after action
})
// Does:
// 1. Quality check screen accessibility
// 2. Find element by name/label
// 3. Tap coordinates
// 4. Optionally type text
// 5. Optionally take verification screenshot
// Returns: { success: true, tappedElement: {...}, screenshot?: {...} }Cost: ~90 tokens (vs 130 tokens separately) Latency: ~300ms (vs ~400ms separately) Use case: User login, form submission, navigation flows
Performs complete app refresh: shutdown → (erase) → boot → build → install → launch
workflow-fresh-install({
projectPath: "./MyApp.xcworkspace",
scheme: "MyApp",
simulatorUdid: "...", // optional: auto-detects
eraseSimulator: true, // optional: wipe simulator data
configuration: "Debug",
launchArguments: ["--resetData"]
})
// Does:
// 1. Shutdown simulator if running
// 2. Erase simulator state (if requested)
// 3. Boot simulator fresh
// 4. Build app
// 5. Install app
// 6. Launch app with arguments
// Returns: { success: true, buildTime: 7000, bootTime: 3000, launchTime: 500 }Cost: ~200 tokens (vs 300+ tokens separately) Latency: ~20s (vs 25+ seconds separately) Use case: CI/CD pipelines, clean state testing, fresh debugging sessions
simctl-device — Simulator lifecycle (7 operations)
boot,shutdown,create,delete,erase,clone,rename- Auto-UDID detection, performance tracking, smart defaults
simctl-app — App management (4 operations)
install,uninstall,launch,terminate- Bundle ID resolution, launch arguments, environment variables
idb-app — IDB app operations (4 operations)
install,uninstall,launch,terminate- Physical device + simulator support via IDB
cache — Cache management (4 operations)
get-stats,get-config,set-config,clear- Multi-layer caching (simulator, project, response, build settings)
persistence — Persistence control (3 operations)
enable,disable,status- File-based cache across server restarts
idb-targets — Target management (2 operations)
list,describe,connect,disconnect- Physical device and simulator discovery
Build & Test (6 tools)
xcodebuild-build: Build with progressive disclosure via buildIdxcodebuild-test: Test with filtering, test plans, cache IDsxcodebuild-clean: Clean build artifactsxcodebuild-list: List targets/schemes with smart cachingxcodebuild-version: Get Xcode and SDK versionsxcodebuild-get-details: Access cached build/test logs
UI Automation (6 tools)
idb-ui-describe: Accessibility tree queries (all | point operations)idb-ui-tap: Coordinate-based tapping with percentage conversionidb-ui-input: Text input with keyboard controlidb-ui-gesture: Swipes, pinches, rotations with coordinate transformsidb-ui-find-element: Semantic element search (NEW in v2.0)accessibility-quality-check: Rapid UI richness assessment (NEW in v2.0)
I/O & Media (2 tools)
simctl-io: Screenshots and video recording with semantic namingscreenshot: Vision-optimized base64 screenshots (inline, max 800px)
Discovery & Health (3 tools)
simctl-list: Progressive disclosure simulator listing (96% token reduction)simctl-get-details: On-demand full simulator data retrievalsimctl-health-check: Xcode environment validation
Utilities (5 tools)
simctl-openurl: Open URLs and deep linkssimctl-get-app-container: Get app container paths (bundle, data, group)simctl-push: Simulate push notificationsrtfm: On-demand comprehensive documentation
Workflow Tools (2 high-level abstractions) - NEW in V3.0.0
workflow-tap-element: High-level semantic tap (find + tap in one call)workflow-fresh-install: Clean install workflow (shutdown → erase → boot → build → install → launch)
Total: 29 active tools (27 core + 2 workflow abstractions)
// 1. Quality check before choosing approach
accessibility-quality-check({ screenContext: "LoginScreen" })
// → { quality: "rich", tappableElements: 12, textFields: 2 }
// 2. Find email field semantically
idb-ui-find-element({ query: "email" })
// → { centerX: 200, centerY: 150, label: "Email", type: "TextField" }
// 3. Tap and input email
idb-ui-tap({ x: 200, y: 150 })
idb-ui-input({ operation: "text", text: "user@example.com" })
// 4. Find and tap login button
idb-ui-find-element({ query: "login" })
// → { centerX: 200, centerY: 400, label: "Login", type: "Button" }
idb-ui-tap({ x: 200, y: 400 })
// 5. Verify (screenshot only for confirmation, not primary interaction)
screenshot({ screenName: "HomeScreen", state: "LoggedIn" })Efficiency Comparison:
- Accessibility approach: 4 queries × 50 tokens = 200 tokens, ~500ms total
- Screenshot approach: 3 screenshots × 170 tokens = 510 tokens, ~6000ms total
- Savings: 2.5x cheaper, 12x faster
// 1. Browse tool categories
rtfm({ categoryName: "build" })
// Returns:
{
category: "build",
tools: [
{ name: "xcodebuild-build", description: "Build Xcode projects with smart defaults" },
{ name: "xcodebuild-test", description: "Run tests with filtering and test plans" },
...
]
}
// 2. Get comprehensive docs for specific tool
rtfm({ toolName: "xcodebuild-build" })
// Returns:
{
tool: "xcodebuild-build",
description: "Full comprehensive documentation...",
parameters: { projectPath: "...", scheme: "...", configuration: "..." },
examples: [...],
relatedTools: ["xcodebuild-clean", "xcodebuild-get-details"]
}
// 3. Execute with discovered parameters
xcodebuild-build({
projectPath: "./MyApp.xcworkspace",
scheme: "MyApp",
configuration: "Debug"
})// 1. Build returns summary + buildId
xcodebuild-build({
projectPath: "./MyApp.xcworkspace",
scheme: "MyApp"
})
// Returns:
{
buildId: "build-abc123",
success: true,
summary: {
duration: 7075,
errorCount: 0,
warningCount: 1,
configuration: "Debug",
sdk: "iphonesimulator"
},
nextSteps: [
"Build completed successfully",
"Use 'xcodebuild-get-details' with buildId for full logs"
]
}
// 2. Access full logs only when debugging
xcodebuild-get-details({
buildId: "build-abc123",
detailType: "full-log",
maxLines: 100
})
// Returns: Full compiler output, warnings, errorsCopy this into your project's CLAUDE.md to guide AI agents toward optimal XC-MCP usage:
# XC-MCP Optimal Usage Patterns
This project uses XC-MCP for iOS development automation. Follow these patterns for maximum efficiency.
## Tool Discovery
1. **Browse categories**: `rtfm({ categoryName: "build" })` — See all build-related tools
2. **Get tool docs**: `rtfm({ toolName: "xcodebuild-build" })` — Comprehensive documentation
3. **Execute**: Use discovered parameters and operations
## Accessibility-First Automation (MANDATORY)
**ALWAYS assess accessibility quality before taking screenshots:**
1. **Check quality**: `accessibility-quality-check({ screenContext: "LoginScreen" })`
- Returns: `rich` | `moderate` | `minimal`
2. **Decision branch**:
- IF `rich` or `moderate`: Use `idb-ui-find-element` + `idb-ui-tap` (faster, cheaper)
- IF `minimal`: Fall back to `screenshot` (last resort)
3. **Why this matters**:
- Accessibility: 50 tokens, 120ms per query
- Screenshots: 170 tokens, 2000ms per capture
- **3-4x cheaper, 16x faster when accessibility sufficient**
- **Promotes inclusive app development**
## Progressive Disclosure
- Build/test tools return `buildId` or cache IDs
- Use `xcodebuild-get-details` or `simctl-get-details` to drill down
- **Never request full logs upfront** — get summaries first
## Best Practices
- **Let UDID auto-detect** — Don't prompt user for simulator UDIDs
- **Use semantic context** — Include `screenContext`, `appName`, `screenName` parameters
- **Prefer accessibility over screenshots** — Better for efficiency AND app quality
- **Use operation enums** — `simctl-device({ operation: "boot" })` instead of separate tools
## Example: Optimal Login Flow
\`\`\`typescript
// 1. Quality check (30 tokens, 80ms)
accessibility-quality-check({ screenContext: "LoginScreen" })
// 2. IF rich: Semantic search (40 tokens, 120ms)
idb-ui-find-element({ query: "email" })
idb-ui-tap({ x: 200, y: 150 })
idb-ui-input({ operation: "text", text: "user@example.com" })
idb-ui-find-element({ query: "login" })
idb-ui-tap({ x: 200, y: 400 })
// 3. Verify with screenshot only at end (170 tokens, 2000ms)
screenshot({ screenName: "HomeScreen", state: "LoggedIn" })
// Total: ~280 tokens, ~2400ms
// vs Screenshot-first: ~510 tokens, ~6000ms (2.5x slower, 1.8x more expensive)
\`\`\`- macOS with Xcode command-line tools
- Node.js 18+
- Xcode 15+ recommended
Install Xcode CLI tools:
xcode-select --install# Global install (recommended for MCP)
npm install -g xc-mcp
# Or run directly without installation
npx -y xc-mcp
# Local development
git clone https://github.com/conorluddy/xc-mcp.git
cd xc-mcp && npm install && npm run buildClaude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):
{
"mcpServers": {
"xc-mcp": {
"command": "npx",
"args": ["-y", "xc-mcp"],
"cwd": "/path/to/your/ios/project"
}
}
}Environment Variables (optional):
XCODE_CLI_MCP_TIMEOUT: Operation timeout in seconds (default: 300)XCODE_CLI_MCP_LOG_LEVEL: Logging verbosity (debug | info | warn | error)XCODE_CLI_MCP_CACHE_DIR: Custom cache directory pathXC_MCP_DEFER_LOADING: Enable deferred tool loading (default: true for V3.0)
What Changed:
- All 29 tools now have
defer_loading: trueflag - Claude's platform tool search discovers tools automatically
- No custom tool-search implementation needed
- Tools loaded on-demand based on conversation context
Migration Path:
| Scenario | Action | Notes |
|---|---|---|
| New Projects | No action needed | Platform handles discovery |
| Existing Integrations | No action needed | Compatible with V2.x usage |
| Debugging/Testing | Set env var | Use XC_MCP_DEFER_LOADING=false |
Usage (same as V2.x):
// V3.0 - Platform discovers tools automatically
// Just use tools as before - Claude's tool search handles discovery
xcodebuild-build({ scheme: "MyApp" })
// Use RTFM for documentation discovery
rtfm({ categoryName: "build" })
rtfm({ toolName: "xcodebuild-build" })
// Disable defer_loading for debugging
export XC_MCP_DEFER_LOADING=falseToken Impact:
| Version | Startup | Discovery | Notes |
|---|---|---|---|
| V2.0.x | ~18.7k | N/A | All tools loaded upfront |
| V3.0.0 | ~0 | Platform-managed | Tools loaded on-demand |
npm run build # Compile TypeScript to JavaScript
npm run dev # Development mode with watch compilation
npm test # Run Jest test suite (60% coverage)
npm run test:coverage # Generate coverage report
npm run lint # ESLint with auto-fix
npm run format # Prettier code formatting- Jest with ESM support and TypeScript compilation
- 60% coverage across statements, branches, functions, lines
- 1136 tests covering core functionality, edge cases, error handling
- Pre-commit hooks enforce code quality via Husky + lint-staged
Core Components:
src/index.ts— MCP server with tool registration and routingsrc/tools/— 29 tools organized by category (xcodebuild, simctl, idb, cache, workflows)src/state/— Multi-layer intelligent caching (simulator, project, response, build settings)src/utils/— Shared utilities (command execution, validation, error formatting)src/types/— TypeScript definitions for Xcode data structures
Cache Architecture:
- Simulator Cache: 1-hour retention, usage tracking, performance metrics
- Project Cache: Remembers successful build configurations per project
- Build Settings Cache: Auto-discovers bundle IDs, deployment targets, capabilities
- Response Cache: 30-minute retention for progressive disclosure
Warning
I appreciate contributions, but please note that this repo and my other public repos are far down in the priority queue of what I'm working on, so I'll be slow to review anything. Your best bet is really just to fork the repo and customise it to your own needs.
PR requirements:
- Tests pass (
npm test) - Coverage remains ≥60% (
npm run test:coverage) - Code passes linting (
npm run lint) - TypeScript compiles (
npm run build)
See CLAUDE.md for detailed development guidelines and architecture documentation.
MIT License — See LICENSE for details.
XC-MCP: Production-grade Xcode automation for AI agents through progressive disclosure and accessibility-first workflows.