TextMorph
A system-wide, privacy-first multimodal AI writing assistant & floating toolbar for Android. Powered by a modular multi-engine architecture capable of executing any open foundation LLM locally with zero-copy memory mapping, slash commands, and an embedded OpenAI server.
On-Device AI in Action
Explore TextMorph's official Google Play Store visuals highlighting on-device LiteRT-LM & llama.cpp inference, floating accessibility overlay bar, slash commands, and embedded localhost OpenAI server.
System-Wide Floating AI Agent
Universal AccessibilityService & WindowManager Overlay Across Any App
A living, intelligent floating companion widget anchored over any Android application. Powered by a non-intrusive AccessibilityService and WindowManager overlay (FLAG_NOT_FOCUSABLE), it lets you trigger autonomous agent workflows, voice missions, or instant screen analysis over WhatsApp, YouTube, Chrome, or Settings without switching apps.
Key Capabilities
- Living companion bubble with expressive visual status cues (thinking, acting, completed)
- Universal overlay functioning over any 3rd-party Android app without keyboard focus loss
- Single-tap screen assist & long-press quick autonomous agent mission trigger
- Non-intrusive draggable pill with smooth edge snapping and animated pulse halos
AccessibilityService · WindowManager Overlay · FLAG_NOT_FOCUSABLE · Multi-Step Dispatcher All Play Store Cards
Key Features
Multi-Engine On-Device AI
Modular multi-engine inference runtime (Google LiteRT-LM and native engines) capable of executing any open foundation LLM directly on mobile silicon with hardware GPU/NPU acceleration and total privacy.
System-Wide Floating AI Pill
Non-intrusive AccessibilityService overlay that anchors above active input fields without stealing keyboard focus (FLAG_NOT_FOCUSABLE).
Zero-Copy Downloads Loader
Directly memory-maps multi-gigabyte .litertlm model files via Linux mmap and ParcelFileDescriptor with zero duplicate flash storage overhead.
Universal Slash Commands
Trigger AI transformations in any text box by typing /fix, /summarize, /fr (French), or /ar (Arabic) with instant in-place buffer replacement.
Embedded Localhost OpenAI Server
Lightweight Ktor CIO HTTP server on 127.0.0.1:8080 exposing standard OpenAI-compatible streaming chat completions for Termux and local tools.
Multimodal Vision Screen Assist
MediaProjection integration captures active screen context and provides instant on-device multimodal reasoning and answers without cloud exposure.
Direct Companion Actions
Top 5 customizable prompt actions accessible with a single tap in the floating companion bar. Drag-and-drop reordering with persistent Room DB.
Zero-Permission Action Menu
Registers directly as an Android ACTION_PROCESS_TEXT context menu provider for highlight-and-transform operations in any browser or app.
Charcoal & Slate Modern Material Design
Bespoke aesthetic balancing deep charcoal slate tones with high-contrast neutral accents and adaptive full-width dialog properties.
Architecture & Technology
Jetpack Compose M3 + Floating Companion WindowManager Bespoke Charcoal Slate palette with living floating companion bubble anchored via WindowManager (FLAG_NOT_FOCUSABLE), full-width responsive dialogs, and tactile cards.
Code-as-Action + Tri-Modal Perception + Magnetic Snapping Autonomous multi-step device task executor with deepest-leaf text matching, magnetic target snapping, click_element(id), precision reticle overlays, and SkillMemoryService.
Sherpa-ONNX KWS + 3-Attempt Acoustic Calibration 100% offline keyword spotting for custom wake-words ("Computer", "Hey Tamim") with 3-attempt acoustic tuning studio, hardware hotkeys, and automated voice script runners.
LiteRT-LM Gemma 4 Vision + Zero-Copy mmap + llama.cpp Hardware-accelerated mobile GPU inference supporting Google Gemma 4 (with Vision tokens) and Gemma 3 1B with zero-copy Linux mmap direct from /sdcard/Download.
Ktor CIO + Server-Sent Events (127.0.0.1:8080) Asynchronous non-blocking HTTP server listening on 127.0.0.1:8080. Implements standard OpenAI /v1/models and streaming /v1/chat/completions for local edge automations.
Live CPU/RAM/Thermal Profiler + Zero Process Death Contract Real-time on-device hardware profiler monitoring LiteRT-LM CPU load, RAM footprint (5.7GB), and thermals (38°C), paired with near-logging (AppLogger) and zero-crash contracts.