autoresearch

What it does

Turns Claude Code, OpenCode or Codex into a loop that keeps iterating until a machine-checkable number improves — constraint plus metric plus autonomy, following Karpathy's formulation. Removing the human from each judgement is exactly what makes it compound, and exactly where it goes wrong: a badly chosen metric gets optimised at the expense of the thing you actually wanted, and an unattended loop bills continuously. Set the metric carefully and cap the spend. MIT.

Summary based on the project's own description. The Japanese page has a fuller beginner guide: setup difficulty, what to prepare, and a step-by-step start.

Screenshots & visuals

Before you start

• Setup difficultyEasy to start
• Command lineRequired
• API keyNot required
• Sends data onlineYes

Judged automatically from the project's own metadata. Always confirm on the official page.

autoresearch FAQ

Is autoresearch free?

Yes. It runs on your own machine, so there is no usage-based cost either.

Do I need to install anything or write code to use autoresearch?

It needs the command line. If you are not comfortable typing commands, expect to get stuck during setup.

Does autoresearch work on a phone?

No. It needs a computer.

Can I use autoresearch commercially?

The licence is MIT, which permits commercial use. Conditions such as keeping the copyright notice still apply, so read the licence text before shipping it inside something.

How it compares

The conditions that decide whether you can actually use it, side by side.

"Sends data out" means your input is uploaded to someone's server. Pick "No" when the material cannot leave your organisation.

Other claude code ecosystem tools