vNext MCP Analysis Tools

Visual overview

Use the workflow to follow the task, and the architecture map to separate responsibilities. These are conceptual maps; the guide below defines implementation details and verification limits.

1. 1Define a supported rule↓
2. 2Specify evidence and limits↓
3. 3Implement focused checks↓
4. 4Verify false-positive cases

1. Boundary 1Rule contract
2. Boundary 2Analyzer implementation
3. Boundary 3Finding schema

Connected responsibilities, not a required class hierarchy or an execution trace.

Context

Use this as the contract for expanding ios-agent-mcp beyond the current Swift engineering analyzers. These tools remain static, read-only project reviewers. Runtime behavior belongs in ios-simulator-mcp.

Implemented in the 2.6.0 source

review_app_intents: SiriKit migration advice plus opt-in schema and onscreen entity association checks. See integration contract. Parameter-summary/localization/identifier analysis remains future work.

Tool Plan

Tool	Focus	Typical signals
review_ui_ux	Visual structure, native conventions, state coverage	hardcoded spacing, fixed fonts, missing loading/empty/error states, inconsistent radii, excessive GeometryReader
review_motion	SwiftUI animation correctness	broad .animation, missing value dependency, infinite animation, layout-heavy animation, Reduce Motion gaps
review_accessibility	Semantic and inclusive design	missing labels, small tap targets, fixed sizes, contrast-risk colors, ignored Dynamic Type
review_haptics	Meaningful feedback	haptics fired on appear, duplicated feedback, no reduced-sensory fallback, heavy impact for minor state
review_realitykit	3D scene safety and native fit	unbounded asset loads, missing collision/input components, ARKit without permission copy, no fallback
review_metal	GPU rendering safety	per-frame allocation, missing drawable guard, unsafe buffer sizing, no pixel-format/depth consistency
review_webkit	Web/native interop safety	untyped JS bridge, broad navigation, injected secrets, WKWebView used where native controls fit better
review_foundation_models	Foundation Models usage	missing availability gates, no graceful fallback, unsafe prompt logging, absent evaluation path
review_core_ai	Core AI model integration	model lifecycle, privacy boundaries, device capability checks, background work isolation

| review_ai_security | AI privacy and misuse resistance | prompt injection surfaces, secret leakage, unbounded tool calls, unsafe retrieval context | | review_ai_evaluations | Evaluation coverage | no datasets, no code-based evaluators, no regression gate, no failure taxonomy | | review_networking | Network correctness and resilience | unbounded retries, no cancellation, live API defaults in previews/tests, missing offline state | | review_persistence | Data and storage safety | UserDefaults secrets, main-actor I/O, missing migrations, model objects crossing actor boundaries | | review_storekit | Purchase readiness | missing restore path, unverified transactions, no pending/refund handling, live StoreKit in tests | | review_permissions | Permission and entitlement correctness | missing purpose strings, overbroad entitlements, no denied/restricted state, privacy manifest gaps |

Output Contract

Every new review tool should match the existing analyzer shape:

{
  "summary": "Motion review found 3 issues.",
  "score": 82,
  "counts": {
    "critical": 0,
    "high": 1,
    "medium": 2,
    "low": 0
  },
  "files_checked": 18,
  "issues": [
    {
      "rule": "animation-without-reduced-motion",
      "severity": "high",
      "file": "Sources/App/HomeView.swift",
      "line": 42,
      "message": "Animated transition has no Reduce Motion alternative.",
      "why": "Users who disable motion can still receive large movement.",
      "fix": "Read accessibilityReduceMotion and switch to opacity or no animation."
    }
  ],
  "suggestions": [
    "Centralize animation tokens for spring duration and response."
  ]
}

Rule Design

Prefer high-signal rules over style opinions:

• Flag literal spacing only when repeated enough to indicate no token system, or when it creates inconsistent layout.
• Flag fixed font sizes when used in user-facing text without a Dynamic Type path.
• Flag animation problems when they can cause incorrect behavior, inaccessible motion, or performance cost.
• Flag AI issues when they expose privacy, availability, evaluation, or tool-call safety risk.
• Do not score "premium design" from static code alone. Leave aesthetic confirmation to the Visual Iteration Loop.

Static vs. Runtime Boundary

Static analyzer can say	Runtime loop must prove
A view uses fixed font sizes	Text actually fits at accessibility sizes
A button appears to lack a label	VoiceOver announces the correct label and trait
A Metal renderer allocates per frame	GPU frame pacing is stable on the simulator/device
A splash view ignores Reduce Motion	The captured intro avoids large motion when Reduce Motion is enabled
A Foundation Models flow lacks evaluations	The model meets pass/fail thresholds on a dataset

Anti-Patterns

// WRONG: add a review tool whose finding is only "looks bad".
Why: static analyzers need reproducible evidence.

// RIGHT: detect concrete code patterns, then route aesthetic verification to screenshots.

// WRONG: call App Intents or model providers during static review.
Why: ios-agent-mcp is read-only and network-free.

// RIGHT: inspect declarations, availability gates, privacy handling, and test/evaluation files.

// WRONG: create separate output shapes per tool.
Why: clients cannot branch reliably.

// RIGHT: reuse summary, score, counts, files_checked, issues, and suggestions.