Quick Start

Quick Start Guide

Option 1: Chat (No GPU, No Model Download)

git clone https://github.com/itallstartedwithaidea/MiniAgent.git
cd MiniAgent
pip install -r requirements.txt
python -m miniagent.chat

This runs in knowledge-base mode. Ask about CPA, ROAS, GAQL, campaign structure, negative keywords, impression share.

Option 2: Web Demo

pip install streamlit
streamlit run scripts/web_demo.py

Option 3: Install MCP Server (Connect to Real Google Ads)

pip install -e ".[google]"
claude mcp add miniagent-google -- python -m miniagent.mcp.google_ads

Then in Claude Code: "Analyze my Google Ads campaigns"

Option 4: Install Agent Skills

/plugin marketplace add itallstartedwithaidea/miniagent
/plugin install advertising-full@miniagent

Option 5: Train Your Own Model (2 Hours, 1 GPU)

git clone https://github.com/itallstartedwithaidea/MiniAgent.git
cd MiniAgent
make download-data
make train
make chat

Or step by step:

pip install -r requirements.txt
python scripts/download_data.py --all
python trainer/pretrain.py --dim 512 --n_layers 8
python trainer/sft.py --load_from ./checkpoints/pretrain_512.pth
python -m miniagent.chat --model ./checkpoints/sft_512.pth

Option 6: Run Benchmarks

make eval
# or
python eval/advertising_bench.py --quick

Option 7: Run Quickstart Example

python examples/quickstart.py

This runs: create model → forward pass → generation → benchmarks → chat test — all in one script.