OpenResearch is a local-first desktop app and command-line workspace for running many AI research agents side by side, each one confined to its own agent session and its own isolated git worktree. It is built for people who do experimental work with code — machine-learning researchers, graduate students, and engineers who want to try ten variants of an idea at once instead of babysitting a single thread. The project comes from alphaXiv, is written mostly in Rust, and is released under the MIT license.
What it does
The pitch is parallelism plus a paper trail. Most agent tooling runs one experiment, with one model, in one working directory; OpenResearch is designed so that each research direction gets an independent agent session and a separate worktree, so runs do not collide and nothing gets overwritten mid-flight.
Around that it adds the bookkeeping that experimental work usually loses:
- Parallel exploration. One agent session and one isolated worktree per direction.
- Reproducible experiments. Variants are tracked in a git-native experiment tree, and every run receives an immutable archive of the commit it was recorded against.
- Evidence in context. Logs, diffs, files, results and artifacts stay attached to the work that produced them, rather than scattering across terminals and scratch folders.
- Your choice of agent. Claude Code, Codex, OpenCode and Cursor are all supported, and you can pick a different one per experiment.
The README also frames the workspace as a home for "autoresearch": agents that review literature, develop hypotheses, run experiments and produce research artifacts. The channel's write-up goes a step further and describes OpenResearch proposing, running and iterating on experiments on its own, closing the loop between an idea and a result.
How it works
Git does the isolation. Because each agent works in its own worktree off the same repository, parallel sessions share history without sharing a checkout, and the experiment tree is just git structure rather than a separate database you have to trust. The immutable archive per run is what makes a result auditable later: you can point at a specific commit and say this is what produced that number.
Execution is local by default. The CLI, orx, brings up a dashboard on http://127.0.0.1:4791, and your code, logs and results stay on your machine. Beyond the local box, the materials describe running over SSH on a remote machine, or on managed OpenResearch compute when you want more hardware than you own — the same workspace, a different place for the work to happen.
Model choice is deliberately open. Alongside the hosted coding agents, the documentation covers connecting LM Studio, oMLX, Ollama or a custom endpoint through OpenCode, so a fully local model stack is a supported path rather than an afterthought.
Getting started
There are two front doors. The desktop app is a download from the project's site, and on macOS or Linux the CLI installs with a single shell command followed by orx up, which opens the local dashboard. Windows is handled through the GitHub Releases page; the project marks that support as still in beta, requires Git for Windows, and asks you to read its Windows notes first.
An account on the project's site is optional for local use — it exists for email updates and for the managed compute option. Everything else runs without one.
When to use it / when not
Reach for OpenResearch when your bottleneck is breadth rather than a single long run: sweeps over prompts or hyperparameters, several competing hypotheses you want to explore at the same time, or a codebase where you want agents making changes without stepping on each other. The reproducibility story matters most when someone — a reviewer, a collaborator, or you in three months — has to reconstruct which version of the code produced which result.
It is a worse fit if your work is not in a git repository, since the whole isolation model rests on worktrees. It is also early: the repository was created in June 2026 and has been pushed to as recently as September 2026, with around 1,400 stars and just under a hundred forks — real interest, but not a decade of hardening. Windows users in particular should expect rough edges. And running many agents in parallel multiplies model costs, so the appeal of the local-model path is practical as well as philosophical.
If you already work with a coding agent and keep hitting the limit that it can only chase one idea at a time, OpenResearch is worth an afternoon. The combination of per-experiment worktrees, an immutable record tied to a commit, a free choice of agent per run, and local-first execution addresses a specific and familiar failure mode: a pile of half-remembered experiments nobody can reproduce. Researchers and engineers doing exploratory work on their own hardware are the audience that will feel the difference; anyone who runs one careful experiment at a time can safely watch this one mature first.