OpenResearch

Projects & Experiments

Agents do research inside projects. Every experiment they try, and every run of it, is recorded on your computer so later work can build on earlier results.

Projects

A project is a local git repository. Create one from the app's New Project form:

  • Blank project: a fresh repository.
  • Existing folder: a repository you already have.
  • From a paper: search alphaXiv and clone the paper's public code, or start blank with the PDF, then ask the agent to reproduce it.

Syncing the project to GitHub is optional, under Advanced.

Experiments

Experiments form a tree. Each one has a title and description that state its hypothesis, a run command, and its own git branch (orx/<slug>). Root experiments are baselines; a child branches from its parent and inherits its run command. Once a run has answered an experiment, it is frozen, and new ideas become new children.

Runs

A run executes an experiment's committed branch on a compute backend; uncommitted changes are not included. On OpenResearch, Hugging Face, Modal, and Kubernetes, runs time out after 4 hours unless you pass --timeout. Runs end as done, failed, cancelled, or lost. Logs, metrics, and artifacts are saved locally.

From the CLI

Agents drive all of this through orx, and you can too:

bash
orx projects
orx project view <projectId>
orx create-experiment <projectId> --title "Lower LR" --parent <expId>
orx exp run <expId> --backend ssh --host runpod-gpu
orx runs <projectId>
orx logs <runId>
orx metrics <runId>       # Trackio metrics
orx tensorboard <runId>   # prints a tensorboard command

Run orx skill for the full guide agents read.

Where data lives

Projects, experiments, and runs are stored in a local SQLite database in ~/.local/share/openresearch by default, along with run logs and artifacts. Change the location in Settings > Storage or with ORX_DATA_DIR. OpenResearch doesn't upload your code, logs, or agent conversations.