Cortex

Cortex — Native AI Agent Manager for Android

A native Android app that turns your device into a multi-agent AI workspace: specialized agents with editable system prompts, persistent per-project workspaces, 17+ LLM providers, on-device voice and local models, and multi-agent delegation orchestrated from a central coordinator.

Highlights

🤖 Multi-agent system

• Create agents with specialized roles (developer, accountant, lawyer, tutor, researcher, …) and fully editable system prompts
• Per-agent temperature and max-token configuration
• A central coordinator delegates work across agents:
  • DELEGATE: [Agent] — single agent
  • DELEGATE_SEQ: [A] -> [B] — sequential chains (output chaining)
  • DELEGATE_PAR: [A], [B] — parallel execution

🗂️ Workspaces

• Separate workspaces per project, each with its own active agent and model
• Persistent conversation history and file attachments per workspace

🔌 17+ LLM providers

Managed Cortex, OpenRouter, Google AI, OpenAI, NVIDIA, Anthropic, Moonshot, MiniMax, DeepSeek, xAI (Grok), Ollama, LM Studio, Kilo, Alibaba, OpenCode, Z.AI — plus on-device local models (MediaPipe, downloaded from HuggingFace) for fully offline inference.

🎙️ Voice, on-device first

• 4 STT backends with automatic hardware detection: Sherpa-ONNX (Snapdragon NPU), MediaPipe Whisper, Vosk, and cloud fallback
• Embedded TTS voices — no voice API keys required
• Invoke Cortex as the Android system assistant role, or use the floating bubble for hands-free conversation

🔒 Security

• API keys stored in EncryptedSharedPreferences (AndroidX Security Crypto)
• Integrated terminal with risk-level command approval before any shell access
• Optional per-tool execution permissions

🧩 Extensible

• MCP server support, persistent memory (FTS + links), scheduled tasks, skills, and subagent executions
• Tool integrations: shell, files, search, calendar, GitHub, Notion, Slack, Google Drive, memory, todos, finance, skills, delegation, app controls

🧠 Knowledge base (on-device RAG)

• Semantic search over your own documents and notes, fully on-device (MediaPipe Text Embedder, Universal Sentence Encoder — multilingual)
• Documents are chunked, embedded, and stored in Room; retrieval is tool-called with mandatory SOURCE citations, and retrieved content is treated as user data, never as instructions
• No API keys, no cloud calls, no data leaves the device; brute-force cosine search is fine for personal-scale collections
• Add documents from Settings → Knowledge base (RAG), then ask Cortex about them in any chat

☁️ Backend + dashboard

• Firebase Cloud Functions backend (Node.js 22) with automated tests
• Web dashboard for monitoring and voice-pack distribution

Architecture

MVVM + Clean Architecture — Jetpack Compose · Hilt · Room (v53, 19 entities) · Retrofit · KSP

• data/ — Room database + migrations, Retrofit clients, repositories, orchestration, tool handlers, STT services
• domain/ — pure domain models (Agent, Workspace, Message, Provider, MCPServer, LocalModel)
• presentation/ — Compose screens + ViewModels, one package per screen (13 modules)
• di/ — Hilt modules

Build

./gradlew assembleDebug          # debug APK
./gradlew installDebug           # install on a connected device
./gradlew test                   # unit tests
./gradlew connectedAndroidTest   # on-device instrumentation tests

A sideload build variant distributes the app outside Google Play together with a hosted STT/TTS voice pack (./gradlew prepareSideloadVoicePack).

Firebase setup

1. Copy app/google-services.json.example → app/google-services.json and fill in your Firebase Android app values (package com.aiagents.app).
2. Never commit the real google-services.json — it is gitignored; CI injects it from the GOOGLE_SERVICES_JSON repository secret.
3. Deploy backend secrets with firebase functions:secrets:set.

CI

Three GitHub Actions workflows: ci (build + unit tests), beta-quality-gate, and android-instrumentation (on-device tests).

Project layout

app/          Android app (Kotlin, Jetpack Compose)
functions/    Firebase Cloud Functions backend (Node.js 22)
dashboard/    Web dashboard
voice/        Voice pack sources
docs/         Design docs and plans
benchmark/    Performance benchmarks

License

MIT © 2026 Gabriel Hernandez