A git-native laboratory for your coding agents

What it does

OpenResearch is a local-first desktop workspace that turns familiar coding agents into autonomous research assistants. It spins up isolated git worktrees for parallel experiments, tracks every run as an immutable archive of its recorded commit, and keeps evidence—logs, diffs, artifacts—tied to the agent session that produced it. The whole loop, from literature search to hypothesis testing, happens on your machine or on remote GPUs you control.

The interesting bit

The project treats reproducibility as a version-control problem: experiments branch like code, and every run is snapshotted to a git commit so you can actually trace which idea spawned which result. It also runs fully offline by default, with cloud accounts only gating managed compute and team features.

Key highlights

• Agent-agnostic harness: Works with Claude Code, Codex, or OpenCode, picking the model per session.
• Git-native experiments: Each experiment lives in its own worktree; the experiment tree preserves lineage across parallel agent explorations.
• Remote-friendly: The same committed snapshot runs over SSH, Slurm, Kubernetes, Ray, and various cloud job runners without publishing the repository.
• Local ownership: Projects, conversations, and artifacts stay in a local SQLite store on loopback by default.
• Autoresearch loop: Agents can propose ideas, edit code, launch experiments, inspect results, and decide what to try next.

Caveats

• Windows support is still in beta and requires Git for Windows.
• Remote workspaces bind to loopback with no application-level authentication, so other users on the same host can reach the dashboard.
• Official release builds send coarse, opt-out usage analytics tied to a random installation ID (source builds do not).

Verdict

Worth a look if you want coding agents to actually run experiments without turning your repo into an untraceable pile of scratch files. Less useful if you already have a lab notebook and cluster workflow you love.