Voice-first interaction
A continuous voice loop with wake-word handling, speech recognition, turn control, and local text-to-speech.
A computer-native
AI operator.
An on-device desktop assistant that combines local model routing, voice interaction, persistent memory, MCP tools, and controlled desktop actions behind a FastAPI service.
RuntimeLocal-first
InterfaceVoice + desktop
Control planeFastAPI
01 / What Myrah can do
One assistant.
Many local systems.
Local model orchestration
Routes requests across local assistant, coding, reasoning, vision, and embedding models instead of relying on one remote model.
Memory that stays useful
Stores conversation context and explicit preferences in a local memory layer, with knowledge ingestion and retrieval workflows.
Actions with boundaries
Exposes desktop, browser, system, productivity, and VS Code tools through MCP with permission-aware safety controls.
02 / The architecture
Private by default.
Capable by design.
PythonFastAPIOllamaMCPSQLitellama.cpp
The runtime combines a FastAPI service, Ollama or llama.cpp model backends, SQLite memory, knowledge ingestion, voice services, and MCP servers for desktop, browser, system, productivity, and VS Code actions.
Health + system APIModel inventoryDesktop actionsSafety evaluation
03 / Runtime modes
A system that
keeps context.
Chat
Local conversational assistant with session context and model routing.
Operator
Natural-language intent becomes a structured desktop action plan.
Memory
Conversation turns and explicit preferences persist in SQLite.
Knowledge
Local Markdown, text, PDF, and EPUB sources can be ingested and searched.
Curated local model bundle
qwen3:4bqwen2.5-coder:3bdeepseek-r1:1.5bqwen2.5vl:3bnomic-embed-textfaster-whisper-large-v3
Myrah routes general, coding, reasoning, vision, embedding, and speech tasks to purpose-fit local models rather than forcing every request through one model.