OpenResearch turns the coding agent you already use — Claude Code, Codex, OpenCode, Cursor or Google Antigravity — into a research agent that reviews papers, proposes hypotheses, runs experiments and stores evidence on your own machine. alphaXiv published v0.2.10 on September 23, 2026, with precompiled binaries for macOS, Linux and Windows (the latter still in Beta).
The proposal isn’t a new agent. It’s a workspace around the one you already pay for and trust with your repo.
What is OpenResearch?
OpenResearch is an open source CLI with a local dashboard, created by alphaXiv, the team behind the discussion platform for arXiv. Its README defines it as “the local-first workspace for research agents and autoresearch”.
Specifically, it adds to your coding agent the structure that research work needs and that a chat session doesn’t have:
- Parallel exploration. Each research direction has its own agent session and its own isolated git worktree.
- Reproducible experiments. Variants are recorded in a git-native experiment tree, and each run receives an immutable file of the commit it ran against.
- Evidence in context. Logs, diffs, results and artifacts remain associated with the work that produced them.
- Agent and model per session. You choose the harness (Claude Code, Codex, OpenCode, Cursor or Antigravity) and the model in each session.
Antigravity is the most recent harness: it arrived with PR #332, already merged, and an approval bypass fix (PR #391) was integrated on September 20.
Why does each agent work in its own git worktree?
Because that way three hypotheses can advance at the same time without stepping on each other’s files. A git worktree is a second working directory hooked to the same repository, and OpenResearch assigns one to each session. It’s what makes parallel research agents possible without one’s output contaminating the other’s.
What is autoresearch?
Autoresearch is OpenResearch’s autonomous mode: the agent proposes an idea, changes the code, launches an experiment, reviews the evidence and decides what to try next. Multiple agents can do this in parallel over different directions, and the experiment tree preserves each one’s lineage. That way you can trace what idea produced what result, instead of reconstructing it from the chat history.
Does it work as AI for scientific research?
Yes, and that’s exactly its territory: experimental research, that is, training, evaluation, benchmarks and ablations. It doesn’t replace a web search assistant. If you’re tuning models, comparing prompts at scale or benchmarking implementations, this tool is for you. If you just want your agent to “read about a library”, it’s more than you need.
Literature also enters the cycle: orx paper <arxiv-id-or-doi> fetches a paper by its arXiv ID or DOI, and orx discover keyword <query> searches by topic.
How do you turn Claude Code into a research agent?
You install the OpenResearch CLI, bring up the workspace and install the skill on your agent. On macOS or Linux:
curl -LsSf https://openresearch.sh/install.sh | sh
orx up
orx up opens the local dashboard at http://127.0.0.1:4791.
Then install the OpenResearch skill on the coding agents you have:
orx install-skills
There’s also a macOS app (macOS 11 or higher, universal for Apple Silicon and Intel) that brings up the dashboard without a terminal. According to the docs, the app doesn’t add orx to your PATH, so install the CLI as well if you want to use the commands.
Does OpenResearch work on Windows?
Yes, but in Beta at the time of publishing this note (September 23, 2026). The Windows version requires Git for Windows to be installed first. The v0.2.10 release page also includes a PowerShell installer:
powershell -ExecutionPolicy Bypass -c "irm https://github.com/alphaXiv/OpenResearch/releases/download/v0.2.10/openresearch-cli-installer.ps1 | iex"
That command is pinned to v0.2.10; check the releases page to see the current version.
Which orx commands will you use?
These are the usual commands listed in the README:
orx projects
orx project view <project-id>
orx runs <project-id>
orx logs <run-id>
orx exp run <experiment-id>
orx discover keyword <query>
orx paper <arxiv-id-or-doi>
orx --help and orx <command> --help show the complete interface.
Can it be used with Ollama or other local models?
Yes, through the OpenCode harness, not Claude Code: OpenCode runs the agent loop, and LM Studio, oMLX (Apple Silicon) or Ollama serve the model on your machine. According to the documentation, you don’t need any cloud account. Any other OpenAI-compatible server (vLLM, for example) works as a custom endpoint.
The connection is made within OpenResearch:
- Go to Settings → Harnesses → OpenCode → Add local model, or choose Add Local Model under OpenCode in the model selector.
- Enter the server address. The documented default values are
http://127.0.0.1:1234/v1for LM Studio,http://127.0.0.1:11434/v1for Ollama, andhttp://127.0.0.1:8000/v1for oMLX. - Click Find models and choose a model that supports tool calling.
The documentation suggests 32K context as a starting point for agent tools.
A warning that the documentation itself makes explicit: local inference doesn’t mean offline. Paper searches, downloads, and GitHub operations still use the network. If you’re still setting up your OpenCode installation, our list of plugins and agents for OpenCode worth installing is a good complement.
Is OpenResearch free?
The local workspace doesn’t require an account. According to the README, the openresearch.sh account is only used for service functions, such as organizations and managed compute. The documentation presents openresearch.sh as a compute marketplace, where you compare GPU and CPU availability and pricing across providers and launch instances from the web or with orx.
The model is clear: local use and on your own infrastructure is open, and hosted compute is the paid product.
Does your code leave your machine?
Not by default: OpenResearch runs on 127.0.0.1 with a local SQLite store, and creating a project or launching an execution doesn’t publish your code.
Before adopting it, it’s worth knowing two things.
Telemetry is opt-out. Official builds send general usage events associated with a random installation ID. According to the README, those events exclude code, prompts, paths, repo names, and tokens. It’s disabled with orx telemetry off, checked with orx telemetry status, or omitted per command with --no-telemetry. Builds compiled from source code send nothing.
Remote mode has no authentication. orx up --remote user@host brings up the workspace along with remote GPUs while you use the browser on your laptop. The remote service listens on loopback, but has no application-level authentication, so any other user of that host can access it. On a shared GPU server, that matters.
The same confirmed snapshot can also run via SSH or on Slurm, Kubernetes, Ray, Hugging Face Jobs, Modal, and Tinker, without needing to publish the repository.
What changed in v0.2.10?
We don’t know: at the time of publishing this note, the v0.2.10 tag included no release notes, only installers and binaries for macOS (Apple Silicon and Intel), Windows x64, and Linux x64 and ARM64 (musl). The project is pre-1.0 and publishes versions frequently, so pin a version if you’re going to integrate it into something others depend on.
Is it worth installing?
Yes, if your work revolves around experiments and not features. OpenResearch’s strength is that it doesn’t compete with your agent: it adds what agents lack to do research. That’s isolation between parallel attempts, recording which commit produced which number, and evidence that doesn’t disappear when you close the chat. It’s the most concrete tool of this week on GitHub for getting more out of the agent you already have.