Skip to content

Quick Start

Get AgenticART running in 5 minutes.

Prerequisites

  • Python 3.10+
  • Android emulator (Genymotion or AVD)
  • Ollama for local LLM inference

Installation

1. Clone and Install

git clone https://github.com/GitSolved/AgenticART.git
cd AgenticART

# Install Core & Dojo dependencies
pip install -r requirements.txt
pip install -r dojo/requirements.txt

# For M3 Max (Fine-tuning support)
pip install mlx-lm

2. Start Android Emulator

# For Android Studio AVD (Rooted recommended)
emulator -avd <your_avd_name>

3. Start Ollama

ollama serve

# Recommended (32B Coder) - Requires 24GB+ RAM
ollama pull qwen2.5-coder:32b

# Fast Alternative (7B Coder) - Requires 8GB RAM
ollama pull qwen2.5-coder:7b

The RAG system provides contextual knowledge to reduce hallucinations:

# Install RAG dependencies
pip install sentence-transformers chromadb

# Populate knowledge bases with OWASP/CWE data
python scripts/populate_rag.py

This loads: - OWASP Mobile Top 10 2024 vulnerability data - CWE definitions relevant to Android security - Curriculum examples from completed challenges

Run Challenges

Basic Test Run

# Run white belt challenges using Ollama
python -m dojo.test_end_to_end --mode live --belt white --model qwen2.5-coder:32b

This will:

  1. Load White Belt (Observation) challenges.
  2. Send artifacts (APK/Manifest) to the LLM.
  3. Use the Praxis Loop to verify reasoning against the emulator.
  4. Grade the results and capture DPO Training Data.

Execution Modes

Mode Flag Description
Live --mode live Real-time execution via Ollama
MLX --mode mlx Native Apple Silicon execution (Fastest)
Mock --mode mock CI/CD testing without a real LLM

Belt Progression

# Progress through belts as the model improves
python -m dojo.test_end_to_end --mode live --belt white
python -m dojo.test_end_to_end --mode live --belt yellow
python -m dojo.test_end_to_end --mode live --belt orange

Package Training Data

After running challenges, package the data for fine-tuning:

python -m dojo.finetune.packager

This creates training datasets from:

  • ✅ Working scripts
  • ✅ Error→fix pairs
  • ✅ Graded trajectories

Web Interface

Launch the Streamlit dashboard:

streamlit run webapp/app.py

Access at http://localhost:8501

Next Steps