What is OpenResearch?
OpenResearch is a desktop app and CLI (orx) from alphaXiv, the arXiv discussion site, that wraps the coding agent you already use in the tooling a research project needs. Agents can search alphaXiv, bioRxiv and PubMed to ground a hypothesis, launch experiments on your machine or remote compute, and record every run, with its code, logs and artefacts, in a local SQL database the agent consults before deciding what to try next.
It works with Claude Code, Codex, OpenCode, Cursor and Google Antigravity, and with local models through OpenCode and LM Studio, oMLX or Ollama. Experiments can run locally or over SSH, Slurm, Kubernetes, Ray, Modal, Tinker, Hugging Face Jobs or alphaXiv's managed GPUs, which are billed at cost from a prepaid balance.
The repository was created on 7 June 2026, is written in Rust, is MIT-licensed and has about 7,700 GitHub stars. Releases are frequent; v0.2.23 shipped on 31 October 2026. The Windows app is in beta.
Is OpenResearch worth it?
OpenResearch's useful idea is that a research agent needs a lab notebook more than a smarter model. Every experiment goes into a local database with its code, logs and artefacts, so the agent can see what was tried, what failed and what the numbers were before proposing the next run, and a person can inspect the evidence behind each result in the desktop app. That addresses the commonest failure of autoresearch tools: repeating an experiment that already failed, or reporting a result nobody can trace.
The breadth of compute backends is the second reason to look. A researcher can move the same project from a laptop to a university Slurm cluster, a Kubernetes namespace, Modal or Hugging Face Jobs without changing tools, and the literature side draws on alphaXiv, bioRxiv and PubMed rather than a general web search. Because it drives Claude Code, Codex or OpenCode rather than shipping its own agent, it improves as those agents do.
The defaults are less private than the marketing line suggests. Projects, code and traces stay local, but official builds send opt-out product analytics tied to a random install ID, plus agent-written feedback reports on bugs and frustrations that the agent is merely instructed to sanitise. Anonymised token and message counts feed a public model-usage ranking on openresearch.sh. orx telemetry off, or a switch in Settings, disables all of it. The standalone CLI installer is unsigned, so managed Macs may block it.
At under five months old, with about 109 open issues and pull requests and twenty-plus patch releases in the 0.2 line, this is early software. Managed GPU prices are visible only after login. Compared with Sakana's AI Scientist, which runs a fixed pipeline end to end, OpenResearch is a workspace a researcher steers, which makes it more useful day to day and less of a demonstration.
Where it earns its place
A PhD student with a Slurm allocation at a university cluster
Run an ablation sweep and keep track of what each run showed
Connect the cluster over SSH and Slurm, have Claude Code propose and launch runs, and let OpenResearch record each one with logs and figures. The software is free; compute is the existing allocation.
A research engineer at a 15-person ML start-up
Reproduce a new paper's result before deciding whether to build on it
Start a project from the paper, let the agent pull related work from alphaXiv and run the training code on Modal or managed GPUs from a prepaid balance. The experiment log makes a failed reproduction a documented result rather than a lost afternoon.
A computational biologist at a hospital research unit
Ground an analysis plan in recent preprints and published work
Use the bioRxiv and PubMed search through the agent, run analyses locally so patient-adjacent data stays on the machine, and turn analytics off with orx telemetry off before starting.
Before you commit
- Setup
- Download the macOS 11+ app, the Windows beta (needs Git for Windows) or a Linux AppImage (glibc 2.35+), pick and sign in to a coding agent, then start from an empty repository, an existing folder or a paper. The CLI installs with a curl script and orx up opens a dashboard on 127.0.0.1:4791.
- Integrations
- Claude Code, Codex, OpenCode, Cursor and Antigravity; LM Studio, oMLX and Ollama through OpenCode; alphaXiv, bioRxiv and PubMed; local, SSH, Slurm, Kubernetes, Ray, Modal, Tinker, Hugging Face Jobs and managed compute.
- Data and privacy
- Projects, conversations, experiments and artefacts stay local. Official builds send opt-out analytics tied to a random install ID and agent-sanitised feedback reports; orx telemetry off disables both. An account is needed only for managed compute and email updates. MIT licence.
- Cost at scale
- No paid software tier. Managed GPUs are billed at cost from a prepaid balance with rates shown in the dashboard; otherwise cost is your own cluster or cloud bill plus your coding agent's subscription or tokens.
- Limits
- About 109 open issues and pull requests; frequent patch releases (v0.2.23 by 31 October); unsigned CLI may be blocked on managed Macs; managed compute pricing not published publicly.
What’s good
- Local experiment database the agent reads before each run
- Local, SSH, Slurm, Kubernetes, Modal, Ray and HF Jobs
- alphaXiv, bioRxiv and PubMed literature search built in
- Drives Claude Code, Codex, OpenCode, Cursor or Antigravity
What’s not
- Analytics and agent feedback reports on by default
- Managed GPU rates only visible after login
- Windows beta; CLI installer unsigned
- Under five months old and still at 0.2
Best for
ML and computational-science researchers who already use a coding agent and want it to run, log and compare experiments across their own compute.
OpenResearch alternatives
OpenResearch: common questions
- Is OpenResearch worth it?
- OpenResearch gives coding agents the experiment memory and compute plumbing that research actually needs, from a team whose business is papers, but it is young and phones home by default. We rate it 3.8 out of 5.
- How much does OpenResearch cost?
- Free and open source (MIT). Managed OpenResearch compute is billed at cost from a prepaid balance; rates are shown in the dashboard, not published on the site. Agent and model costs are your own.
- What are the downsides of OpenResearch?
- Analytics and agent feedback reports on by default; managed GPU rates only visible after login; windows beta; CLI installer unsigned; under five months old and still at 0.2.
- Who is OpenResearch best for?
- ML and computational-science researchers who already use a coding agent and want it to run, log and compare experiments across their own compute.
- How much does OpenResearch cost for a team or in production?
- No paid software tier. Managed GPUs are billed at cost from a prepaid balance with rates shown in the dashboard; otherwise cost is your own cluster or cloud bill plus your coding agent's subscription or tokens.
- Is OpenResearch safe for company data?
- Projects, conversations, experiments and artefacts stay local. Official builds send opt-out analytics tied to a random install ID and agent-sanitised feedback reports; orx telemetry off disables both. An account is needed only for managed compute and email updates. MIT licence.
- What are the best OpenResearch alternatives?
- The closest alternatives we have reviewed are The AI Scientist (v1, v2 and v3), Scientific Agent Skills and ShinkaEvolve.
Answers checked Oct 31, 2026.