CoAutoResearch

An AI research partner you stay in control of.

An open-source research agent that does the work with you — running trials, drafting your manuscript, and tracking every claim — while you stay able to understand, steer, and defend the project at every step.

Most "auto research" tools optimize for autonomous output: generate ideas, run experiments, write a paper, done. CoAutoResearch optimizes for research ownership — the result isn't a finished artifact you can't explain, it's a traceable research trajectory you can actually use, revise, and defend.

• 0.1.1 — Default dashboard project folder is now co-autoresearch-projects/ (legacy local-projects/ still discovered).
• 0.1.0 — Initial release: immutable project template; co-auto-research CLI (init / ui / doctor / ls / attach); remote-server tunneling; product-docs site.

Full history in CHANGELOG.md.

Quick start

npx --yes co-auto-research ui

This downloads the CLI if needed, starts a local dashboard, and opens it in your browser. Click + in the sidebar to create a project, then describe where your research should begin. Attach local files with the composer + button — they're copied into your project's resources/user_input/attachments/ so the agent can cite them later.

Bringing a proposal, a deep-research report, or a half-finished project? See Best practices for the recommended ways to start — and how to turn the blueprint into a finished paper.

You'll need Node 18+ and one agent CLI — Codex (default) or Claude Code. The few-minute setup is in Getting started.

For repeated use, install it once:

npm install -g co-auto-research
co-auto-research ui

Update anytime with npm install -g co-auto-research@latest.

Closed the terminal? Reopen any project from the same folder:

co-auto-research ls
co-auto-research attach my-project

On a remote server, co-auto-research ui --remote prints an SSH tunnel command so you can open the UI from your local browser — see Remote servers.

Why it's different

• You're the PI. Step in anytime to change direction, scope, methods, claims, or venue — and your decisions become part of the project record.
• Progress you can follow. The agent picks one next objective, plans it, runs it, and reports — no hidden search tree to reverse-engineer.
• Every claim is traceable. Findings, evidence, limitations, and rejected paths are kept separate, not buried in chat logs or generated prose.
• Built for revision. Answer reviewers, defend assumptions, and keep the project moving after the AI hands off.
• Your work stays yours. Each project is an independent copy; the reusable template is never touched during normal work.

What a project looks like

Generated projects keep the pieces that make research usable, separate and legible:

• PROJECT.md — your canonical research direction.
• research_trajectory/ — current state, findings, and a trial-by-trial audit trail.
• resources/ — papers, data, prior work, and target-venue materials.
• workspace/ — the live, executable working area.
• manuscript/ — blueprint, figure specs, reviews, and deliverables.

Documentation

• Getting started — install, prerequisites, first project
• Concepts — how the research loop works
• Best practices — how to start well, and turn the blueprint into a paper
• CLI reference — every command and flag
• Multiple projects · Remote servers · Platform support · Upgrading projects

Contributing

Contributions are welcome — see CONTRIBUTING.md, which also covers running from a source checkout.

Citation

If you use CoAutoResearch in your research, please cite it:

@software{coautoresearch2026,
  title        = {CoAutoResearch: An AI research partner you stay in control of},
  author       = {Tang, Yihong},
  year         = {2026},
  howpublished = {\url{https://github.com/YihongT/CoAutoResearch}}
}

License & contact

Released under the Apache-2.0 license. Questions, feedback, or collaboration: [email protected]