ai-agent

autoresearch: Auto‑run nano‑chat research on a single GPU

Name: autoresearch: Auto‑run nano‑chat research on a single GPU
Author: karpathy

Empower your AI experiments with autoresearch, which trains nano‑chat models automatically on one GPU. Ideal for founders needing fast, hands‑off research. 97k+ GitHub stars.

From the repo · “AI agents running research on single-GPU nanochat training automatically”

98,335 stars+413★ this week13,608 forksPythonGuide quality 8/10 · reviewed Sep 29Updated 3/26/2026100% free · open source

What it does

Autoresearch generates automated research papers and articles using AI agents, allowing startup founders to quickly explore and summarize large amounts of information on a given topic.

When to use it

•When you need to quickly research and summarize a large amount of information on a new topic
•When you want to automate the process of generating research papers or articles
•When you need to explore and analyze large datasets and generate insights

Ready-to-paste prompt

python generate.py --topic 'AI for startup founders' --dataset 'startup_data.csv' --output 'research_paper.md'

Heads up: Make sure you have the correct version of Python (3.8 or later) and the required dependencies installed, as specified in the `requirements.txt` file, to avoid compatibility issues

Saves to your device

Use with Claude

New

Skip the builder — one click puts this in Claude, Cursor, Antigravity and more.

✅ Light setup

Installs with a command or two; your AI agent can do it for you.

Try it instantly — no install

Claude Code

mkdir -p ~/.claude/skills/autoresearch && curl -fsSL https://workflowstacks.com/api/skills/autoresearch/claude-skill -o ~/.claude/skills/autoresearch/SKILL.md

Open in another AI app

Opens the app with this repo with the prompt ready to go — no copy-paste needed.

Connect the whole catalog (MCP)

claude mcp add --transport http workflowstacks https://workflowstacks.com/api/mcp

Adds a WorkflowStacks connector to Claude Code: search and load any skill here by chatting.

How autoresearch: Auto‑run nano‑chat research on a single GPU works

Codeflow

Free to inspect

Autoresearch: Auto‑run nano‑chat research on a single GPU is a small Python project (~1.1k lines across 2 code files). Setup is light: install Python, run one command. Last commit 5 months ago, no license file, no tests found.

Size

Small codebase

~1.1k lines · 2 code files · 9 min skim

Setup

Light setup

Install Python, run one command.

Runs on

Python

No API keys detected

Python 83%Jupyter Notebook 17%

What happens, step by step

AI agent experiments with code

You give

program instructions

1Prepare Data

download training data

prepare.py

2Train Model

train model for 5 minutes

train.py

3Run Experiment

run experiment and check results

4Edit Code

agent edits train.py file

train.py

5Repeat Experiment

repeat experiment with new code

You get

experiment results

Where to start reading

1
README.md
Start here — what it does and how to install it
2
pyproject.toml
Dependencies and the commands it exposes

READMENo tests foundThin docsNo CINo licenseUpdated 5 mo ago

Use with Claude

New

Skip the builder — one click puts this in Claude, Cursor, Antigravity and more.

✅ Light setup

Installs with a command or two; your AI agent can do it for you.

Try it instantly — no install

Claude Code

mkdir -p ~/.claude/skills/autoresearch && curl -fsSL https://workflowstacks.com/api/skills/autoresearch/claude-skill -o ~/.claude/skills/autoresearch/SKILL.md

Open in another AI app

Opens the app with this repo with the prompt ready to go — no copy-paste needed.

Connect the whole catalog (MCP)

claude mcp add --transport http workflowstacks https://workflowstacks.com/api/mcp

Adds a WorkflowStacks connector to Claude Code: search and load any skill here by chatting.

Quick Actions

Details

Creator

karpathy

Language

Python

Do this with a working workflow

Pages we wrote and tested that put a skill like this to work — importable n8n templates, outcome guides and Claude Desktop configs.

Outcome guide

Triage your inbox automatically

The workflow watches your inbox and classifies each new email by priority with a one-line reason and a draft reply, logging everything to a ranked sheet. You open your day to a triaged list, not a pile.

Outcome guide

Reduce meeting no-shows automatically

Keep your meetings in a Google Sheet. Every morning the workflow emails a warm, confirming reminder to anyone scheduled for the next day — automatically, in your voice.