Reproduce the paper baseline
1 run
An alphaXiv project
Autoresearch on your machine
OpenResearch is the local-first workspace for research agents. Start the local dashboard, then give each research direction its own agent. They work in parallel, in isolated worktrees, using the tools and compute you already have.
macOS 11+ and Linux via the CLI. Windows beta uses the same CLI with Git for Windows.
long-context-1bgit-native experiment tree
Test a longer context window
3 runs
Sweep optimizer and schedule
8 runs
Tighten the evaluation harness
4 runs
Ablate the regularizer
5 runs
The workspace
One dashboard for the whole research loop
Describe the question in plain language. OpenResearch opens a session, picks the harness, records the commit and keeps the evidence attached to the run.
harness: Codex model: gpt-5 compute: runpod A100
RunComputeDurationStatus
sweep-optimizer-08
A100 80GB2h 14mrunningsweep-optimizer-07
A100 80GB2h 06mrunninglong-context-03
H100 80GB3h 41mdoneharness-tighten-04
Local M3 Max18mdoneExperiment gallery
Directions an agent can carry on its own
Each row below is a real research direction: an idea, a code change, a run, and the evidence that decided what happened next.
Reproduce the paper baseline
Runs stay attached to the variant that produced them, so the comparison is available long after the session is closed.
Test a longer context window
Runs stay attached to the variant that produced them, so the comparison is available long after the session is closed.
Sweep optimizer and schedule
Runs stay attached to the variant that produced them, so the comparison is available long after the session is closed.
Tighten the evaluation harness
Runs stay attached to the variant that produced them, so the comparison is available long after the session is closed.
Ablate the regularizer
Runs stay attached to the variant that produced them, so the comparison is available long after the session is closed.
Start your own direction
Describe the question, choose the harness and the compute, and let the loop run overnight.
Capabilities
Built for research agents, not demos
The parts that make an agent useful on real research work: isolation, lineage, evidence and control over where it runs.
Parallel exploration
Give each research direction an independent agent session and an isolated git worktree, so two ideas never fight over the same checkout.
Reproducible experiments
Track variants in a git-native experiment tree. Every run receives an immutable archive of the exact commit it recorded.
Evidence in context
Logs, diffs, files, results and artifacts stay tied to the work that produced them, instead of being scattered across folders.
Your choice of agent
Run Claude Code, Codex, OpenCode or Cursor. The harness and the model are selected per session, not per project.
Your choice of compute
Start on your laptop, move to your own infrastructure, or spin up managed GPU capacity from the built-in compute marketplace.
Local ownership
Projects, conversations, experiments, runs, logs, code and artifacts stay on your machine in a local store. Nothing is published by default.
Workflow
From install to a running experiment in four steps
01
Install and start
One command installs the CLI and launches the local dashboard on 127.0.0.1:4791.
02
Point an agent at a direction
Install the OpenResearch skill into your coding agent and describe the question you want answered.
03
Let the loop run
The agent proposes an idea, changes the code, launches the run, reads the evidence and decides what to try next.
04
Keep the record
Every variant keeps its lineage, so a result can be traced back to the commit and the run that produced it.
Compute
Run the same snapshot anywhere
The committed source snapshot is portable, so a variant can start on your laptop and finish on rented GPUs. Publishing the repository is not required.
| Target | How you run it | Notes |
|---|---|---|
| Local | orx up | Runs on 127.0.0.1 with a local store. The default. |
| Remote host | orx up --remote user@host | SSH aliases and custom ports supported. |
| Slurm | Cluster submit | Keep the workspace next to the scheduler. |
| Kubernetes | Cluster submit | Same committed snapshot, cluster execution. |
| Ray | Cluster submit | Distributed trials across the cluster. |
| Hugging Face Jobs | Managed | Offload a run without publishing the repo. |
| Modal | Managed | Serverless GPU jobs per experiment. |
| Tinker | Managed | Managed fine-tuning targets. |
| Compute marketplace | Managed | Runpod, Vast, Lambda, Nebius, DeepInfra, Verda, Vultr and MassedCompute offerings in one place. |
Use cases
What researchers run through OpenResearch
Reproduce a paper
Start from the reference implementation, pin the commit, and turn each deviation into a tracked variant with its own evidence.
Hyperparameter sweeps
Queue the search space, let agents prune the boring branches, and keep the schedule and optimizer that actually helped.
Evaluation harness work
Change the harness without losing the earlier numbers: each harness revision keeps its own runs and diffs.
Long-context and training runs
Move a variant from your laptop to remote GPUs while the browser stays pointed at the same workspace.
Plans
Research plans from $9 a month
Monthly and annual plans bundle experiment credits, managed compute access and workspace seats. Start on your own machine first, then add capacity when a sweep needs it. Cancel any time.
Questions
Frequently asked questions
Where does OpenResearch run?
On your machine. The dashboard binds to 127.0.0.1 and keeps projects, conversations, experiments, runs, logs, code and artifacts in a local store. Creating a project or launching a run does not publish your code.
Which coding agents are supported?
Claude Code, Codex, OpenCode and Cursor. Run orx install-skills to install the OpenResearch skill into a supported agent, then pick the harness and model per session.
How do I get more compute?
Run locally, over SSH, or submit the same committed snapshot to Slurm, Kubernetes, Ray, Hugging Face Jobs, Modal, Tinker, or managed OpenResearch compute, which aggregates offerings from multiple providers.
Do I need an account?
Not for local work. An openresearch.sh account is only used for service-owned capabilities such as organizations and managed compute.
What data leaves my machine?
Official release builds send opt-out, coarse usage events tied to a random installation ID. They exclude code, prompts, file contents and paths, repository names, tokens, emails, and project or experiment identifiers. Turn it off with orx telemetry off.
What does a plan include?
Each plan bundles monthly experiment credits, managed compute access, and the workspace features described above. Annual billing is discounted, and you can cancel whenever you like.
Run your first autoresearch loop tonight
Install the CLI, start the local dashboard, sign in with Google and point your first agent at a question you have been meaning to test.
Runs locally. Code and artifacts stay on your machine.