SilbercueSwift

SilbercueSwift

The fastest, most complete MCP server for iOS development. One Swift binary, 58 tools, zero dependencies. SilbercueSwift has the most complete toolset of any alternative out there.

Built for Claude Code, Cursor, and any MCP-compatible AI agent.

Looking for an alternative to existing iOS MCP servers? SilbercueSwift covers the full feature set of XcodeBuildMCP, Appium-MCP, and iosef in a single binary — plus xcresult parsing, UI automation, code coverage, and up to 75x faster screenshots. See comparison below.

Why SilbercueSwift?

Every iOS MCP server has the same problem: raw xcodebuild output is useless for AI agents. 500 lines of build log, stderr noise mistaken for errors, no structured test results. Agents waste minutes parsing what a human sees in seconds.

SilbercueSwift fixes this. It parses .xcresult bundles — the same structured data Xcode uses internally — and returns exactly what the agent needs: pass/fail counts, failure messages with file:line, code coverage per file, and failure screenshots.

Where SilbercueSwift really shines

Screenshots up to 75x faster — ~316ms ( ~15ms)

Free tier screenshots (~316ms) are faster than XcodeBuildMCP (~1127ms) and most alternatives. Pro brings latency down to ~15ms — 5x faster than Appium, 75x faster than XcodeBuildMCP. Agents can take screenshots freely without penalty at either tier.

Structured test results from xcresult bundles — zero guesswork on failures

When a test fails, the agent gets the error message, the exact file:line, a screenshot of the failure state, and optionally the console output — all parsed from Apple's .xcresult format. No guessing from 500 lines of xcodebuild stderr. This is the difference between "agent knows what broke" and "agent guesses what broke".

Single binary, zero dependencies — 58 tools, install in 10 seconds

brew install silbercueswift — done. 8.5MB native Swift binary. No Node.js, no npm, no Appium server, no Python, no Java, no Rust toolchain. Cold start in ~50ms. The fastest way to get an iOS MCP server running.

Agent reads only what matters — 90% fewer tokens, zero wasted calls (topic filtering )

Free tier already strips noise: 15 known noise processes are excluded at capture time, and duplicate lines are collapsed (79% I/O reduction). Pro adds topic filtering — read_logs categorizes lines into 8 topics and shows only app + crashes by default, with a menu: network(87) lifecycle(12) springboard(8). The agent opens specific topics in one call — no guessing, no iteration.

One call to dismiss all permission dialogs — 3 alerts in 1 roundtrip

Every app shows 2–3 permission dialogs on first launch. Other servers require the agent to screenshot → find button → click, per dialog. handle_alert(action: "accept_all") clears them all in a single call, searching across SpringBoard, ContactsUI, and the active app. Free tier handles alerts individually with accept / dismiss.

Drag & drop with element IDs — 1 call instead of 3

"Drag item A above item B" is a single call: drag_and_drop(source_element: "el-0", target_element: "el-1"). The competition only supports raw coordinates, forcing the agent to find both elements, extract their frames, and build a W3C Actions sequence — 3 calls minimum.

Auto-scroll to off-screen elements — no more manual swipe loops

find_element(using: "accessibility id", value: "Save", scroll: true) scrolls automatically until the element appears. SmartScroll handles UIKit, SwiftUI, and lazy-loaded lists — no guessing scroll direction.

View hierarchy in ~31ms (Free) / ~5ms (Pro) — up to 188x faster element inspection

get_source returns the full UI tree in ~31ms (Free) or ~5ms (Pro). The fastest competitor takes 44ms, most take 250ms+. This makes element inspection practically free for agents.

Navigate in one call — find + tap + settle + screenshot in ~380ms

navigate(to: "Settings") finds the element, taps it, waits for the screen to settle, and returns a verification screenshot — all in a single call. No competitor offers this. Agents save 3-4 tool calls per navigation step.

Batch UI automation — run_plan executes multi-step plans with adaptive decisions

run_plan takes a sequence of UI steps and executes them server-side. When a step needs a decision (unexpected dialog, element not found), it falls back through 4 tiers — from MCP sampling to pause & resume. No more "one tool call per tap" overhead.

Quick Start

Install via Homebrew

brew tap silbercue/silbercue
brew install silbercueswift

Or build from source

git clone https://github.com/silbercue/SilbercueSwift.git
cd SilbercueSwift
swift build -c release
cp .build/release/SilbercueSwift /usr/local/bin/

Configure in Claude Code

One command — installs globally for all projects:

claude mcp add --scope user SilbercueSwift /opt/homebrew/bin/SilbercueSwift

Note: Use the full path (/opt/homebrew/bin/SilbercueSwift). Claude Code starts MCP servers without a full shell PATH, so bare command names won't be found.

Uninstall

claude mcp remove --scope user SilbercueSwift
brew uninstall silbercueswift
brew untap silbercue/silbercue

Free vs Pro

SilbercueSwift ships 49 tools for free — build, test, simulate, automate UI, capture logs, and take screenshots. No time limit, no signup.

Pro adds 9 tools and faster internals for teams and power users who need the full picture.

Pro costs 12 EUR/month. Get a license on Polar.sh, then:

silbercueswift activate <YOUR-LICENSE-KEY>

58 Tools in 14 Categories

Build (5 tools)

Testing & Diagnostics (4 tools)

Simulator (12 tools)

UI Automation (16 tools)

Native input for gestures, WDA for element queries and alerts — no Appium, no Node.js, no Python.

handle_alert — the smartest alert handler

# Accept a single alert with smart defaults
handle_alert(action: "accept")

# Dismiss with a specific button label
handle_alert(action: "dismiss", button_label: "Not Now")

# Batch-accept ALL alerts after app launch (unique to SilbercueSwift)
handle_alert(action: "accept_all")

3-tier alert search — finds alerts across:

1. Springboard — system permission dialogs (Location, Camera, Tracking)
2. ContactsUI — iOS 18+ Contacts "Limited Access" dialog (separate process)
3. Active app — in-app UIAlertController dialogs

Smart defaults — knows which button to tap:

• Accept: "Allow" → "Allow While Using App" → "OK" → "Continue" → last button
• Dismiss: "Don't Allow" (handles Unicode U+2019) → "Cancel" → "Not Now" → first button

Batch mode — accept_all / dismiss_all loops through multiple sequential alerts server-side. One HTTP roundtrip instead of N. Returns details of every handled alert.

These capabilities go beyond what other iOS MCP servers currently offer.

Screenshots (1 tool)

Logs (4 tools)

Smart Log Filtering — 4 layers, zero config

# Start capture (default: smart mode — broad stream, topic filtering enabled)
start_log_capture()

# Read logs — default shows only app logs + crashes + topic menu
read_logs()
# → --- 230 buffered, 42 shown [app, crashes] ---
# → Topics: app(35) crashes(2) | network(87) lifecycle(12) springboard(8) widgets(0) background(3) system(83)
# → Hint: include=["network"] to add SSL/TLS + background transfer logs
# → ---
# → [42 filtered lines]

# Agent sees network(87) and wants SSL details — one call:
read_logs(include: ["network"])

# Narrow stream for production monitoring:
start_log_capture(mode: "app")

# Bypass mode logic with explicit predicate:
start_log_capture(subsystem: "com.apple.SwiftUI")

4 filter layers:

1. Stream-side noise exclusion — 15 known noise processes + subsystem/category exclusions removed before buffering. Server-side filtering in logd — 79% I/O reduction.
2. 3 capture modes — smart (default, broad stream for topic filtering), app (tight, auto-detected bundle ID + process name), verbose (unfiltered).
3. Read-time topic filtering — read_logs categorizes every buffered line into 8 topics (app, crashes, network, lifecycle, springboard, widgets, background, system). Default shows only app + crashes. Agent adds topics as needed — stateless per call.
4. Buffer deduplication — 60 identical heartbeat lines become 2: the line itself + ... repeated 59x.

8 topics with LLM-optimized menu:| Topic | Matches | Use case ||---|---|---|| app (always on) | subsystem == bundleId OR process == appName | Your app: os_log, print(), NSLog() || crashes (always on) | fault-level logs | Crashes from any process || network | trustd, nsurlsessiond | SSL/TLS certs, background transfers || lifecycle | runningboardd, com.apple.runningboard.* | Jetsam, memory pressure, app kills || springboard | SpringBoard | Push notifications, app state || widgets | chronod | WidgetKit timeline, refresh budget || background | com.apple.xpc.activity.* | BGTaskScheduler, background fetch || system | everything else | WARNING: high volume |

Console (3 tools)

Git (5 tools)

Visual Regression (2 tools)

Multi-Device (1 tool)

Accessibility (1 tool)

Localization (1 tool)

Automation (2 tools)

Session (1 tool)

xcresult Parsing — The Killer Feature

The Problem

Every Xcode MCP server returns raw xcodebuild output. For a test run, that's 500+ lines of noise. AI agents can't reliably extract which tests failed and why.

The Solution

SilbercueSwift uses xcresulttool to parse the .xcresult bundle — the same structured data Xcode's Test Navigator uses.

# One call, structured result
test_sim(project: "MyApp.xcodeproj", scheme: "MyApp")

→ Tests FAILED in 15.2s
  12 total, 10 passed, 2 FAILED
  FAIL: Login shows error message
    LoginTests.swift:47: XCTAssertTrue failed
  FAIL: Profile image loads
    ProfileTests.swift:112: Expected non-nil value

  Failure screenshots (2):
    /tmp/ss-attachments/LoginTests_failure.png
    /tmp/ss-attachments/ProfileTests_failure.png

  Device: iPhone 16 Pro (18.2)
  xcresult: /tmp/ss-test-1774607917.xcresult

The agent gets:

• Pass/fail counts — immediate overview
• Failure messages with file:line — actionable
• Failure screenshots — visual context (Claude is multimodal)
• xcresult path — reusable for test_failures or test_coverage

Deep Failure Analysis

test_failures(xcresult_path: "/tmp/ss-test-*.xcresult", include_console: true)

→ FAIL: Login shows error message [LoginTests/testErrorMessage()]
    LoginTests.swift:47: XCTAssertTrue failed
    Screenshot: /tmp/ss-attachments/LoginTests_failure.png
    Console:
      [LoginService] Network timeout after 5.0s
      [LoginService] Retrying with fallback URL...
      ✘ Test "Login shows error message" failed after 6.2s

Code Coverage

test_coverage(project: "MyApp.xcodeproj", scheme: "MyApp", min_coverage: 80)

→ Overall coverage: 72.3%

  Target: MyApp.app (74.1%)
      0.0% AnalyticsService.swift
     45.2% LoginViewModel.swift
     67.8% ProfileManager.swift

  Target: MyAppTests.xctest (62.0%)
     ...

Benchmarks

Measured on M3 MacBook Pro, iOS 26.4 Simulator. All values are median of 5 runs after 2 warmups.

Comparison with other MCP servers

See feature comparison table above for a detailed breakdown vs XcodeBuildMCP, Appium-MCP, and iosef. All three are excellent projects that pioneered iOS MCP tooling. SilbercueSwift combines all their feature sets with deeper integration into a single native binary. The only trade-off: SilbercueSwift is iOS-only (no Android, watchOS, tvOS, or visionOS).

Architecture

SilbercueSwift (8.5MB Swift binary)
├── MCP SDK (modelcontextprotocol/swift-sdk)
├── StdioTransport (JSON-RPC)
└── 58 Tools in 14 Categories
    Build · Test · Simulator · Screenshot · UI Automation
    Logs · Console · Visual Regression · Multi-Device
    Accessibility · Localization · Automation · Git · Session

No Node.js. No Python. No Appium server. No Selenium. One binary.

Requirements

• macOS 13+
• Xcode 15+ (for xcresulttool and simctl)
• Swift 6.0+ (for building from source)
• WebDriverAgent installed on simulator (for UI automation tools)

License

The core binary and all 49 free tools are MIT licensed — see LICENSE. Use them however you want, commercially or otherwise.

Pro tools (9 additional tools + faster internals) require a paid license. The license validation code (LicenseManager.swift) is included in the source for transparency — you can see exactly what it checks and when.

Contributing

Issues and pull requests welcome. See CONTRIBUTING.md for guidelines.