Show HN: AndroJack – A grounding gate for Android AI as…

Trust in AI-generated code is falling because models are structurally incapable of verifying their own training data. In Android development, where the ecosystem moves exceptionally fast, AI memory is often twelve to eighteen months out of date.AndroJack is a…

In 2025, the Stack Overflow Developer Survey asked 49,000 developers about their experience with AI coding tools. The results should alarm every Android engineer: The gap between usage and trust is not a coincidence. It is the product of a structural problem: AI models predict tokens, not APIs. They were trained on a snapshot of the world and have no mechanism to know what changed at API 30, what shipped at Google I/O 2025, or what Google Play now rejects at review time.

For Android developers, this failure mode is uniquely dangerous. Android has the fastest-moving ecosystem in mobile development — a new Compose BOM every month, Navigation 3 going stable after seven years of Nav2, Android 16 rewriting the rules on screen orientation locking — and most AI tools have training data that is six months to two years stale by the time you use them. These are not hypothetical risks.

They are documented failure modes from real developer projects. A published case study from Atomic Robot documented a live Navigation 2 → Navigation 3 migration using both Gemini and Claude — with internet access enabled on both. The conclusion, verbatim: "LLMs still hallucinate versions. Even with internet access, both agents wanted to use an outdated release candidate instead of the stable 1.0.0 release." Navigation 3 went stable in November 2025 after seven years of the same library.

It is a complete architectural rethink: back stacks are now plain Kotlin lists, the monolithic nav graph is gone, and NavDisplay replaces NavController. Google's own migration guide is so aware that AI tools get this wrong that it now contains special "AI Agent:" annotations — instructions embedded directly in the official docs for AI tools to follow. An AI tool that generates Nav2 code for a new Compose project in 2026 is not making a small mistake.

It is creating an architectural incoherence that requires a full rewrite to fix. Jetpack Compose ships a new BOM every month. Since most models' training cutoffs, these APIs changed: Every one of these compiles. Most run without errors. The bugs surface later in CI flakiness, UI regressions, or Play Store review failures — and the developer has no idea the AI was confidently wrong. Android 16 made a platform-level change affecting every published app: on devices ≥600dp — tablets, foldables, ChromeOS — apps can no longer lock screen orientation or restrict resizability.

Google Play requires API 36 targeting by August 2026. An AI tool generating android:screenOrientation="portrait" or android:resizeableActivity="false" today is generating code that will trigger App Compatibility warnings in Play Console, fail large-screen quality checks, and get apps demoted in Play Store search results. Foldable users spend 14× more on apps than phone-only users. Tablet + phone users spend 9× more.

FlipaClip saw 54% growth in tablet users within four months of going adaptive. Kotlin Multiplatform went mainstream in 2025 — over 900 new KMP libraries published, Room added KMP support, companies now hire specifically for KMP skills. When a developer on a KMP project asks an AI tool to add database support, the AI generates Android-only Room code. It compiles. It runs perfectly on Android. The iOS build fails.

The developer spends hours debugging before realizing the root cause: their AI tool does not know KMP exists. AndroJack is a documentation-grounded Android engineering MCP server. It gives your AI coding assistant 20 specialized tools that fetch live, verified answers from official Android and Kotlin sources — instead of predicting from stale training data. Level 1 is passive. The tools are available but the AI decides when to use them.

An AI building a Compose screen may call architecture_reference but skip material3_expressive — and ship M3E violations silently. Level 2 is active and guaranteed. The androjack_grounding_gate system prompt (registered on the server — instructions below) maps every task type to the correct tool. Building Compose UI? The AI is mandated to call material3_expressive first. Adding a dependency? It must call gradle_dependency_checker.

No exceptions. → For full grounding, always activate Level 2. See Getting the Full Guarantee below. That's the job AndroJack exists to do — and nothing else in the current ecosystem does it for Android specifically. Prompt engineering means writing instructions into a system prompt or a markdown file (agents.md, SKILL.md, CLAUDE.md, .cursorrules, etc.) that tell the AI how to behave. The fundamental limit: Prompt engineering controls how the AI responds.

It cannot control what the AI knows. You are still trusting training-time knowledge. RAG means building a vector database of documents (official docs, changelogs, internal wikis), embedding them, and injecting the most semantically similar chunks into the AI's context window at query time. MCP (Model Context Protocol) is a standardized protocol — not a retrieval technique or a prompt strategy — for connecting AI models to live tools and external systems.

The Grounding Gate is not a clever name. It is a real enforcement mechanism built into how MCP tools are described to the AI client. "REQUIRED FIRST STEP. You MUST call this before generating any Android/Kotlin code." "Always call this before adding or updating any dependency in build.gradle." "Only produce Android code after reviewing the above official sources." Because MCP clients (Claude Desktop, Cursor, Windsurf, etc.) present these tool descriptions to the LLM as part of its context, the model treats them as workflow constraints — not suggestions.

Each tool lists the specific failure mode it prevents — not just what it does, but what breaks when it is absent. All 20 tools are read-only. AndroJack fetches and returns information — it never modifies your project files. Note on Google Antigravity: This is Google's standalone agentic IDE (released Nov 2025 with Gemini 3) — not Firebase Studio, not Project IDX. Those are separate Google products.

UI path: Agent pane → … → MCP Servers → Manage MCP Servers → View raw config ⚠️ Antigravity is a locally-installed desktop IDE (antigravity.google/download). Do not confuse with Firebase Studio (.idx/) or Project IDX — those are different products. Place at .kiro/settings/mcp.json (project) or ~/.kiro/settings/mcp.json (global). Or use the CLI: kiro-cli mcp add –name androjack –command npx –args '-y androjack-mcp' –scope workspace Every other Android MCP server in the public registry (minhalvp/android-mcp-server, CursorTouch/Android-MCP, mobile-mcp by mobile-next) does the same thing: ADB device control.

They tap screens, capture screenshots, send keystrokes, and run UIAutomator2 queries. They are excellent QA automation tools. Not one of them knows what a ViewModel is. Not one can tell you whether a Gradle coordinate is current. Not one can distinguish Navigation 3 from Navigation 2. AndroJack owns the only unclaimed category in the Android MCP ecosystem: documentation-grounded Android engineering guidance — the trust layer that AI coding assistants call when they need to verify whether an API actually exists, whether a dependency version is current, and whether a pattern is still the official recommendation.

Android Studio added MCP support on January 15, 2026. The window between "first Android doc-grounded MCP" and "Google ships one themselves" is the window AndroJack is built to fill. AndroJack-MCP is a collaborative effort between human product vision and AI engineering excellence. Installing the tools alone gives you Level 1 grounding — the AI can use them but decides when. For Level 2 mandatory, automatic grounding on every Android task, load the androjack_grounding_gate system prompt.

It is a set of rules registered on the MCP server itself (accessible via the MCP prompts API). It maps every task type to the correct tool: IDEs that support MCP prompt injection (Kiro, Antigravity, JetBrains AI): Select the androjack_grounding_gate prompt from the MCP prompts list at session start. Without this step, tool invocation depends on the AI's judgment. With it, the AI is mandated to consult the correct tool for every decision — architecture, UI, dependencies, and testing.

Join our GitHub Discussions to connect with other developers, ask questions, and share your ideas! We welcome contributions! Please see our Contributing Guidelines for more details. Built because 35% of Stack Overflow visits in 2025 are developers debugging AI-generated code. AndroJack exists so none of those visits are yours.

Summary

This report covers the latest developments in android. The information presented highlights key changes and updates that are relevant to those following this topic.

---