An agentic Machine Learning Engineer that trains state-of-the-art ML models using Claude Code SDK and Google ADK. This is a very simple implemenation demonstraing the power of Claude Scientific Skills for machine learning.

• Python 3.13 or higher
• uv package manager
• Claude Code installed and authenticated (see installation guide)

Install dependencies using uv:

uv sync

Create a .env file in the karpathy directory with your API keys:

OPENROUTER_API_KEY=your_openrouter_api_key_here
AGENT_MODEL=your_model_name_here

The OPENROUTER_API_KEY is required for the agent to function properly.

This is the same environment variable that will be copied to the sandbox directory so the agents can use any API keys you provide here.

Run the startup script to set up the sandbox and start the ADK web interface:

python start.py

This automatically:

1. Creates a sandbox directory with scientific skills from Claude Scientific Skills
2. Sets up a Python virtual environment with ML packages (PyTorch, transformers, scikit-learn, etc.)
3. Copies your .env file to the sandbox
4. Starts the ADK web interface
5. Navigate to http://localhost:8000 in your browser
6. Select karpathy in the top left under 'Select an agent'
7. All outputs will be in the sandbox directory so continue to monitor that as you converse with the agent

Note: Any files you want the agent to use (datasets, scripts, etc.) should be manually added to the sandbox directory.

Join our K-Dense Slack community to connect with other users, share ideas, and get support:

Join K-Dense Slack Community

This repository is designed to work with the Claude Scientific Skills collection of ready-to-use scientific tools and workflows (link). The start.py setup script creates a sandbox that includes scientific skills from this collection so the karpathy agent can leverage specialized ML libraries and scientific workflows. For full details on the skills themselves, see the upstream repository’s README and documentation here.

To set up the sandbox without starting the web interface:

python -m karpathy.utils

Note: Any files you want the agent to use (datasets, scripts, etc.) should be manually added to the sandbox directory.

To run the ADK web interface manually:

adk web

Then navigate to http://localhost:8000 in your browser.

If you want substantially more powerful ML capabilities through a multi-agentic system, sign up for www.k-dense.ai. Currently in closed beta, launching publicly in December 2025.

• Modal sandbox integration - Choose any type of compute you want
• K-Dense Web features - We might make some features from K-Dense Web available here based on interest