Introduction

louped is a local workbench for research on language models: how they behave, and what they cost to run.

You keep your research questions in a project. Each question is an experiment: a README that states the question and how you will test it, and a script that runs the test. Your coding agent writes and runs the experiments; you read the results in louped's app.

How it is usedCopy Anchor Link

1. Make a project. louped init my-research sets up the folders and connects your coding agent.
2. Ask a question. Tell your agent (Claude Code, Codex, Cursor) what to test. It writes the experiment.
3. Run it. Small runs go on your machine. Large ones go to a Slurm cluster, which pushes the results to your remote. Pull brings them into your app.
4. Read the results. The run page shows every item under every condition, what changed, and out of how many. Edit any text file a run wrote beside its rendering.
5. Share and shape it. Push runs for others to pull, or publish a read-only dashboard. Shift+click any rows, cards or fields to ask your agent about them, or to have it change them; it points back at what it means.

Everything runs on your hardware with open-weight models. No API key is needed.