#7373rd of 80 in Agent platforms
karpathy
ML engineer agent that trains models in a local sandbox using Agent Skills
- Stars
- 1.6k
- License
- MIT
- Last commit
- Aug 2026
- Language
- Python
1,364 on 21 Apr 20261,582 on 10 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Runs an agent built on the Claude Agent SDK and Google ADK that writes and runs ML training code inside a local sandbox directory. Setup creates a Python virtualenv with PyTorch, transformers and scikit-learn, and loads skills from the Scientific Agent Skills collection. You drive it from the ADK web interface and collect outputs from the sandbox folder.
Who it is for: Engineers trying an agent that automates ML experiments with Agent Skills
Strengths
- start.py builds the sandbox, ML virtualenv and skills in one step
- Reuses the open Agent Skills standard via Scientific Agent Skills (130+ skills)
- Outputs land in a plain sandbox directory you can inspect while the agent works
- MIT licensed, small Python codebase
Weaknesses
- Requires an authenticated Claude Code install plus an OpenRouter API key
- Sandbox is a local directory; Modal compute integration is only listed as upcoming
- Datasets and scripts must be copied into the sandbox by hand
- README calls it a simple demo; fuller capabilities point to the paid K-Dense service
What it needs
- Needs Python 3.13+, uv, Claude Code, OpenRouter API key, Google ADK
- Models: OpenRouter models via AGENT_MODEL
- port 8000
Quick answers
Is karpathy open source?
Yes. Its code is on GitHub under the MIT license.
Also in Agent platforms
See all 80| Rank | Project | Score |
|---|---|---|
| 1 | AutoGPTRole-based AI agents coordinated by one lead agent, plus a visual workflow builder | 95 out of 100 |
| 2 | daily_stock_analysisLLM-driven daily analysis of watchlist stocks across six markets, pushed to chat apps | 95 out of 100 |
| 3 | LangflowVisual builder for AI agents and workflows, exposed as API or MCP server | 94 out of 100 |
| #4 | deer-flowAgent harness running sub-agents, memory and sandboxes with pluggable skills | 93 out of 100 |
| #5 | DifyVisual builder for LLM workflows, RAG pipelines and agents, with APIs | 92 out of 100 |