louped 0.7.0

A local workbench for research on LLM behavior and efficiency.

Website • Docs

Keep your research questions about language models in a project. Each question is an experiment with a README (the question and the test) and a script that runs it. Your coding agent writes and runs them. You read the results item by item in louped's app. Everything runs on your machine or your cluster, with open-weight models.

Start

pip install 'louped[server,tracking,interp,agent]'
louped init my-research && cd my-research
louped serve

Open http://127.0.0.1:8000 and launch the example experiment.

How it works

1. Make a project with louped init. It also connects your coding agent.
2. Ask a question. Your agent (Claude Code, Codex, Cursor) writes the experiment.
3. Run it on your machine, or send it to a Slurm cluster. With a remote set, the job pushes its results and louped pull brings them in.
4. Read the results: every item under every condition, what changed, and out of how many. Edit a run's write-up beside its rendering. Move what you no longer need to the trash.
5. Share and shape it. Push runs for others to pull, or publish a read-only dashboard. Shift+click rows, cards or fields to ask your agent about them, or to have it change them. The agent points back at what it means.

Already have results? louped view <folder> opens them read-only.

See the docs for writing experiments, changing models, using your agent and running on a cluster.

Develop

You need just, uv and Node 22 with pnpm.

just bootstrap     # install everything and the git hooks
just check         # lint, types, tests, and the app and site builds
just serve         # the app on :8000

License

Apache 2.0