Warped

Warped brings the power of large language models directly to your Android device. Chat with AI privately — no account required, no cloud dependency, and your data never leaves your phone.

LOCAL AI, NO INTERNET NEEDED
Run models entirely on-device using two inference engines:
• LiteRT-LM — Google's on-device AI runtime. Supports text, image, and audio input. Tool calling with 15+ built-in tools: weather, calculator, alarms, location, device controls, and more. CPU and GPU auto-detection.
• llama.cpp — Run GGUF models locally. Import .gguf files from your device storage.

CONNECT TO REMOTE PROVIDERS
When you need more power, connect to the AI services you already use:
• OpenAI — GPT-4o, GPT-4, and O-series models
• Anthropic — Claude 3.5 Sonnet, Claude 3 Opus, and more
• Ollama — Connect to your local Ollama server on LAN
• LM Studio — Use your LM Studio server as a backend
• Custom — Any OpenAI-compatible API endpoint
All API keys are encrypted with AES-256 via the Android Keystore hardware security module.

HUGGING FACE INTEGRATION
Search thousands of models directly from Hugging Face:
• Browse Staff Picks, GGUF models, and LiteRT-LM models
• Download with pause, resume, and progress tracking
• Background downloads via WorkManager with notifications
• Gated model support with Hugging Face token authentication

POWERFUL CHAT FEATURES
• Real-time streaming — watch tokens appear as the AI generates them
• Reasoning display — collapsible think blocks for DeepSeek R1, QwQ, and other reasoning models
• Image input — attach photos to your conversations
• Audio recording — record voice input for compatible models
• Markdown rendering — formatted responses with code blocks, lists, and tables
• Conversation history — all chats saved locally with auto-generated titles
• Generation controls — temperature, top-P, top-K, repeat penalty, max tokens, context size, threads, and seed

100% PRIVATE AND OFFLINE-FIRST
• No user accounts, no sign-up
• Chat history stored locally in a Room database on your device
• Downloaded models stored in app-private storage
• API keys encrypted with hardware-backed Android Keystore
• No telemetry, no analytics, no crash reporting
• No ads, no tracking, no data collection

TOOL CALLING WITH LITERT-LM
When using LiteRT-LM models, the AI can use real tools on your device:
• Productivity — current time, timers, alarms, reminders
• Information — web search, weather, news, translations
• Device — battery level, app launcher, device info
• Utilities — calculator, unit converter, password generator, clipboard
• Location — GPS position, nearby places, directions
Each tool can be enabled or disabled individually. A token budget calculator helps you manage context usage.

WHAT MODELS CAN I RUN?
• LiteRT-LM models (.litertlm format) — optimized for mobile, runs on CPU or GPU
• GGUF models (.gguf format) — quantized LLMs from Hugging Face (llama.cpp compatible)
• Any public model on Hugging Face — search and download within the app

PERMISSIONS
• Internet — only used when you download models or connect to remote servers
• Microphone — only when you explicitly record audio for a chat message
• Location — only when you ask the AI for location-based information via tools
• Notifications — only for download progress