Codex Mobile - Local AI Agent 앱 분석 (10월 1일)

Description: Codex is a powerful, locally-hosted AI coding assistant for Android. It allows you to run Large Language Models (LLMs) in GGUF format directly on your device, ensuring privacy and offline capability. Codex is specifically designed to act as an agentic coding assistant that can read from and write to your device's file system, creating and modifying code projects seamlessly. Core Functionalities: 1. Local Model Inference: Runs GGUF models directly on the Android device using llama.cpp. 2. File System Integration: The AI can create, read, modify, and delete files in a designated project folder using specific XML commands. 3. Multiple Projects: Manage separate projects, each with its own working directory (supports both internal storage and external Storage Access Framework directories). 4. System Prompt Customization: Define exactly how the AI should behave and what tools it can use. 5. "Thinking" (Chain of Thought): Supports models that can think before responding, capturing their reasoning in expandable UI blocks. 6. Context Attachments: Attach files to your chat messages to provide context to the AI. 7. Background Execution: Keeps the screen on during generation to prevent the OS from killing the process. Tutorial: How to use Codex Step 1: Download a Model Download a GGUF formatted model (e.g., from HuggingFace). We recommend small quantized models (e.g., Q4_K_M or Q5_K_M variants of 1.5B to 8B parameter models) for optimal performance on mobile devices. Step 2: Load the Model 1. Open the Codex app. 2. Navigate to the "Settings" menu or the top bar in a project chat. 3. Tap the "Load GGUF Model" icon (Robot/SmartToy icon). 4. Select the downloaded GGUF file from your device's file manager. The app will copy and load it. Step 3: Create a Project 1. Go to the main "Projects" screen. 2. Tap the "+" (Add) button. 3. Give your project a name and select a directory where the AI will save its files. You can choose a custom folder on your device. Step 4: Start Coding 1. Open your newly created project. 2. In the chat, prompt the AI. For example: "Create a simple HTML clicker game." 3. The AI will generate the response and use the internal ToolExecutor to automatically create the files in your project directory. 4. You can see the result of the tool execution in the chat (e.g., "[Tool Result: File index.html created successfully]"). 5. Use a file manager or code editor on your device to view and run the generated files! Settings Configuration: - Context Size: Adjust the context window. Higher values use more RAM but allow for larger files and longer conversations. - Keep Screen On: Enable this to prevent the device from sleeping during long text generations. - Global System Prompt: Modify the instructions given to the AI. This is where the tool definitions are defined. Do not remove the tool definitions if you want the AI to write files. - Thinking Enabled: Toggle this if you are using a model that outputs [thinking] tags (like DeepSeek-R1 variants).