MobHunt: Claude Code Configuration

CRITICAL RULES

App Acquisition Integrity: NEVER download mobile binaries (IPAs, APKs, XAPKs) from third-party websites or services (APKPure, APKMirror, Aptoide, etc.). ALL app binaries MUST be acquired exclusively from official sources:

• iOS: App Store via ipatool (requires authenticated Apple ID)
• Android: Google Play or Aurora Store on an Android emulator, then pull via ADB (tools/acquisition/android_acquire.sh) Third-party mirrors cannot be trusted; they may contain tampered, outdated, or repackaged binaries. If official store download fails, ask the user to provide the binary manually. NEVER suggest or attempt 3P download as a fallback.

No 3P Credential Tools: NEVER hand user credentials (passwords, tokens, API keys) to third-party CLI tools or libraries beyond minimal crypto primitives. Build our own tools instead. This includes tools like apkeep, even if open-source.

Python Environment: ALWAYS activate the virtual environment at .venv before installing Python dependencies. Create it with python3 -m venv .venv if it doesn't exist.

Project Overview

MobHunt is an agentic mobile security research system for automated vulnerability hunting through bug bounty programs. It uses an orchestrator agent that spawns focused sub-agents, each assigned a narrow objective (one vulnerability class or attack surface area per agent).

See DESIGN.md for full architecture documentation, and docs/ios-decryption.md for the iOS decryption fallback chain.

Workflow

/hunt <program>  →  scope → acquire → recon → hunt → chain → validate → report
                                                 ↑                        |
                                                 └──── agentic loop ──────┘

Rules (Always Active)

The following rules are loaded on every session and must always be followed:

• rules/hunting.md: Mobile hunting discipline (scope verification, 5-min rule, concrete evidence)
• rules/reporting.md: Report quality standards (impact-first titles, mobile context, exact repro steps)
• rules/ethics.md: Responsible disclosure boundaries (in-scope only, no user data, static first, no publication without approval)
• rules/mobile-specific.md: Platform-specific rules (Apple Silicon, decryption verification, FAT binaries)

Agents

Agents are spawned by the orchestrator with specific, narrow objectives:

Model assignments are configurable in config/providers.json. Three cost profiles are available: quality, balanced (default), and economy. Set cost_profile in providers.json to switch.

Commands

Skills

Skills are contextually loaded by agents and commands:

Configuration

1. Copy config/config.example.json → config/config.json and fill in credentials
2. Copy config/providers.example.json → config/providers.json and configure LLM providers
3. Install external dependencies (see DESIGN.md → External Dependencies)

Both config.json and providers.json are gitignored. Credentials belong in those files or in the environment; never pass a password, token, or API key as a command-line argument, because argv is readable by any other process on the machine and lands in shell history.

Key Directories

Runtime output (gitignored, never published)

These four directories do not exist in a fresh checkout. The tool creates them as it runs, and they fill with target-specific material: acquired binaries, decompiled code, extracted secrets, findings, and draft reports. All of it is confidential to the program being tested.

Never commit, push, or publish anything under these four, and never quote their contents into a public document. Paths are configurable under directories in config/config.json; if you repoint them, keep the new locations out of version control too.

Repository content (committed)

External Tools

Required: ipatool, jadx, apktool, Ghidra Optional: Binary Ninja (set analysis.binary_analyzer to "binja" in config), Semgrep, Mariana Trench Pre-installed (macOS/Xcode): otool, codesign, plutil, nm, lipo, strings