Vikas Sahani

Posted on Mar 2 • Edited on Mar 18 • Originally published at androjack-web.netlify.app

I Built an MCP Server to Stop AI from Hallucinating Android Code — Here's What Happened

#android #mcp #ai #kotlin

description: "AI tools are fast. But in the Android ecosystem, fast + wrong = wasted hours. I built AndroJack MCP to force AI assistants to verify official docs before writing a single line of code."
tags: android, mcp, ai, kotlin
cover_image: https://androjack-web.netlify.app/og-image.png

canonical_url: https://androjack-web.netlify.app

Confidence is the most dangerous trait in an AI coding assistant.

It Started With a Healthcare App That Kept Breaking

While developing a complex, offline-first healthcare application, I found myself stuck in a relentless correction loop.

My AI assistants — Cursor, Claude, Windsurf — were exceptionally fast at generating code. But they were consistently generating the wrong code.

❌ Suggested Navigation 2 patterns for a Navigation 3 project
❌ Hallucinated Gradle coordinates that did not exist on Maven
❌ Used APIs removed from the Android platform years ago
❌ Generated AsyncTask — removed in API 33
❌ Generated TestCoroutineDispatcher — removed from coroutines-test 1.8+
❌ Generated ContextualFlowRow — deprecated in Compose 1.8
❌ Added android:screenOrientation="portrait" — illegal on ≥600dp under Android 16

Every session started with promise and ended with me on developer.android.com manually verifying what the AI should have checked itself.

This was not a failure of the models. It was a failure of grounding.

The Trust Crisis Is Real

I started researching the scale of this problem and found I wasn't alone.

The 2025 Stack Overflow Developer Survey (49,000 developers) confirmed a worrying trend:

84%  of developers use AI tools
29%  trust AI output accuracy  (down from 40% the year before)
35%  of Stack Overflow visits are now triggered by debugging AI-generated code

The structural flaw is simple: AI models predict tokens based on stale training data.

In the Android ecosystem — where Jetpack Compose, Navigation, and Material 3 evolve faster than model weights are retrained — memory is a liability, not an asset.

Passive prompts like agents.md or .cursorrules are not enough. They control how an AI responds. They cannot control what the AI actually knows.

What I Built: AndroJack MCP

I built AndroJack to bridge this gap.

It is an open-source Model Context Protocol (MCP) server that gives your AI assistant a live, physical connection to official Android and Kotlin documentation.

Instead of predicting from stale training memory, the AI is forced to fetch evidence before writing a single line.

VIKAS9793 / AndroJack-mcp

AndroJack: AI that actually knows Android. Real-time dependency tracking, modern architectures, and zero hallucinations.

AndroJack — The Jack of All Android Trades

An MCP server that equips your AI coding assistant with live, verified Android knowledge — so it builds from official sources, not from memory.

Also works with: Windsurf · VS Code Copilot · Google Antigravity · JetBrains AI — see Manual Config below ↓

VS Code distribution: Also live on the VS Code Marketplace as AndroJack MCP for VS Code.

PM / APM docs: Product strategy, JTBD, personas, roadmap, user stories, competitive analysis, and GTM materials live under product-management/README.md.

🔥 The Crisis That Created This Tool

In 2025, the Stack Overflow Developer Survey asked…

View on GitHub

🌐 Website → androjack-web.netlify.app
📦 npm → npmjs.com/package/androjack-mcp

See It in Action

How It Works: 21 Specialized Grounding Tools

AndroJack equips your AI assistant with a complete verification toolset:

Tool	Purpose
`android_official_search`	Live search across developer.android.com + kotlinlang.org
`gradle_dependency_checker`	Real-time version lookup from Google Maven + Maven Central
`android_component_status`	Is this API stable, deprecated, or removed?
`architecture_reference`	Official MVVM, UDF, modularization patterns
`android_api_level_check`	API availability by minSdk level
`kotlin_best_practices`	Idiomatic coroutines, Flow, Compose patterns
`material3_expressive`	Full M3 Expressive reference (GA on Android 16)
`android_navigation3_guide`	Navigation 3 (stable Nov 2025) — stops Nav2 hallucinations
`android_api36_compliance`	Android 16 / Play Store August 2026 mandate checker
`android_code_validator`	Loop-back gate — validates every code block before delivery
`android_kmp_guide`	Kotlin Multiplatform patterns
`android_wearos_guide`	Wear OS 5 + M3 Expressive for Wear
`android_xr_guide`	Android XR SDK (DP3, Dec 2025)
`android_ondevice_ai`	AICore + ML Kit Gen AI patterns
`android_large_screen_guide`	Adaptive layouts, WindowSizeClass, foldables
`android_testing_guide`	Unit, Compose UI, Hilt testing patterns
`android_build_and_publish`	R8, signing, Baseline Profiles, Play publishing
`android_scalability_guide`	Paging 3, offline-first, modularization
`android_permission_advisor`	Runtime permissions + ActivityResultContracts
`android_play_policy_advisor`	Play Store policy (October 2025 changes)
`android_debugger`	Stacktrace diagnosis grounded in official sources

The Grounding Gate: Two Levels of Enforcement

Level 1 — The Suggestion

Tools are available. The AI can use them when it feels uncertain.
Helpful, but still leaves room for overconfidence.

Level 2 — The Guarantee ⭐

Using the androjack_grounding_gate system prompt, you mandate a Verify-Before-Write loop. Every task type is mapped to a required tool call:

Before ANY Android or Kotlin code:

1. SEARCH      → android_official_search
2. VALIDATE    → android_component_status  (every API class)
3. ARCHITECTURE → architecture_reference
4. DEPENDENCIES → gradle_dependency_checker (every library)
5. API LEVEL   → android_api_level_check
6. PATTERNS    → kotlin_best_practices
7. DEBUG       → android_debugger          (for stacktraces)
8. VALIDATE OUTPUT → android_code_validator (every code block)

No code reaches the user without passing through the gate.

The Loop-Back Validator (Level 3)

android_code_validator is the final safety net — a 24-rule validation engine that inspects every code block before it reaches the user:

Verdict: PASS / WARN / FAIL

Catches:
❌ AsyncTask              → removed API 33
❌ TestCoroutineDispatcher → removed coroutines-test 1.8+
❌ ContextualFlowRow      → deprecated Compose 1.8
❌ GlobalScope.launch     → coroutine leak
❌ NavController in new code → use Navigation 3
❌ screenOrientation lock → illegal Android 16 / ≥600dp
❌ runBlocking on UI thread → ANR risk
❌ kapt in new code       → use ksp

If verdict is FAIL — the AI fixes all violations and re-runs before the user ever sees the output.

Get Started in 60 Seconds

Install into any MCP-capable IDE with a single command:

npx androjack-mcp@1.6.4 install

The interactive installer auto-detects your IDEs and writes the config:

✔ System scan complete.

IDE Detection Results:
◉  Claude Desktop     detected, not configured
◉  Cursor             detected, not configured
○  Windsurf           not found
◉  VS Code (Copilot)  detected, not configured
◉  Google Antigravity detected, not configured

◆  AndroJack MCP  Installer
◆  Select installation mode:
   ▸ Auto-install to all detected IDEs  [Recommended]
     Pick specific IDEs
     Show manual config snippet

Manual Config (if you prefer)

Claude Desktop / Cursor / Windsurf / JetBrains:

{
  "mcpServers": {
    "androjack": {
      "command": "npx",
      "args": ["-y", "androjack-mcp@1.6.4"]
    }
  }
}

VS Code (/.vscode/mcp.json):

{
  "servers": {
    "androjack": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "androjack-mcp@1.6.4"]
    }
  }
}

What's New in v1.6.4

This release addresses three areas discovered during the first real-world adoption wave:

🔴 CLI Routing Fix (Critical)

The install wizard was unreachable across all invocations — every subcommand fell through silently to the stdio server instead.

# All of these incorrectly started the server in v1.6.1:
npx androjack-mcp install
npx androjack-mcp --help
npx androjack-mcp --version

# All fixed in v1.6.4 ✅

Root cause: index.ts entry point never read process.argv before defaulting to server start. The install and serve handlers existed as separate entry points but were never routed to.

🟡 IDE Detection UX Improvement

Kiro and JetBrains/Android Studio now surface a clearer status when installed but MCP is not yet configured — instead of silently showing not found.

Before: AWS Kiro  not found
After:  AWS Kiro  detected (MCP not yet configured — launch IDE once to initialize)

🟢 Security Hardening

Updated @modelcontextprotocol/sdk to latest
Resolved all npm audit findings
Improved exception handling across tool handlers
Added input validation checks on all tool entry points
Improved dependency hygiene

Security score: 10.0 / 10.0 on MCP Marketplace scanner — zero known CVEs, verified package integrity.

Early Traction (First Launch Cycle)

AndroJack was distributed across four developer channels simultaneously:

📦 npm Registry          → 240 weekly downloads
🧩 VS Code Marketplace   → 59 installs
🔍 MCP Registries        → 115K+ discovery exposure
📊 Registry Ranking      → Top 300 MCP tools, upward movement

Conversion signal: The VS Code Marketplace showed a 163% conversion rate — meaning more installs happened directly from IDEs than from Marketplace page visits. This indicates community-driven word-of-mouth distribution, not search-driven discovery.

The Negative Constraints (What It Refuses to Generate)

One of the most useful features is what AndroJack prevents. These are hard prohibitions enforced at the system prompt level — the AI cannot generate these without explicit user override:

Universal (all targets):

AsyncTask — crashes on API 33+
TestCoroutineDispatcher — breaks CI on coroutines-test 1.8+
GlobalScope.launch / GlobalScope.async — coroutine leaks
ContextualFlowRow / ContextualFlowColumn — deprecated Compose 1.8
startActivityForResult() — use registerForActivityResult()
IntentService — deprecated API 30
runBlocking {} on UI/main thread — ANR

New Compose projects:

NavController / NavHost — use Navigation 3
MutableLiveData — use MutableStateFlow
kapt — use ksp

targetSdk ≥ 36 (Android 16 / Play Store August 2026 mandate):

android:screenOrientation="portrait" — illegal on ≥600dp
android:resizeableActivity="false" — illegal on ≥600dp

Why MCP Is the Right Architecture for This

MCP (Model Context Protocol) — open-sourced by Anthropic in late 2024 — gives AI assistants a standardized way to call external tools and retrieve live data during a conversation.

This is precisely what grounding requires:

Without MCP:
User prompt → Model predicts from training memory → Code output

With AndroJack MCP:
User prompt → Model calls android_official_search
           → Model calls gradle_dependency_checker
           → Model calls android_component_status
           → Model calls android_code_validator
           → Verified code output

The AI becomes an evidence-based engineer, not a pattern predictor.

What's Coming Next

v1.7.0 — Expanded IDE detection (process-level detection, not config-directory-only)
Community tool contributions — PR-based tool additions
HTTP transport mode — for team-shared MCP instances (already in serve subcommand)

Try It

npx androjack-mcp@1.6.4 install

🌐 androjack-web.netlify.app
💻 github.com/VIKAS9793/androjack-mcp
📦 npmjs.com/package/androjack-mcp

If you're an Android developer using AI-assisted workflows, I'd genuinely love your feedback — especially on which hallucinations you're hitting most often. Every real-world bug report shapes the next iteration.

"You are an evidence-based Android engineer, not a pattern predictor. Your value is not in knowing Android — it is in verifying Android before writing a line."

— AndroJack Grounding Gate system prompt

DEV Community