An autonomous research partner that self-improves recursively and works with you.
Define a question. Run experiments. Discuss findings. Refine the next step.
Build a manuscript and paper draft from evidence you can trace.
Quick start → · Features · Example papers · Documentation
🌐 Languages
English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español
Português brasileiro · Français · Deutsch · Русский · العربية
These are README translations. The interface and full documentation are in English; agent replies follow your language.
- 2026-09-11 — CoAutoResearch v2.0.0 released.
Discuss while research runs. Explore ideas in a separate chat. Review and send a suggestion when you want it to guide the research.
Recursive self-improvement. Feed recorded results, review feedback and reusable lessons into the next trial. Refine hypotheses, methods and research decisions within your brief and budget; keep negative findings and stop when the evidence calls for it.
Inspect the evidence. Follow findings, checks and limitations into a manuscript and a paper draft.
Browse a completed Digits project: research records, manuscript and PDF. This edited, cropped browser tour shows existing history, not a new run or actual research speed.
Three ways to begin
A new question: “Compare two methods on a small public dataset. Use CPU only, state the budget, and ask before final evaluation.”
Existing work: “Read my proposal and attached materials. Identify the next useful experiment and prepare a plan for review.”
A result to investigate: “Check whether this finding survives a stronger comparison. Preserve the original results and report limitations.”
Supported coding agents: ✓ Codex CLI · ✓ Claude Code
Give your coding agent this instruction:
Set up and launch CoAutoResearch from
https://github.com/YihongT/CoAutoResearch.
Follow docs/agent-setup.md, reuse my existing coding-agent login,
configure paper generation, verify the setup, and open the dashboard.
Setup has three checkpoints: dashboard opens → coding agent is authenticated and ready → paper tools are ready. Complete interactive provider login yourself. Model availability and usage limits depend on your account; research can require additional dependencies.
Manual setup from source
You need Node.js 20+, Python 3.10+, Git, and an installed, authenticated Codex or Claude Code CLI.
git clone https://github.com/YihongT/CoAutoResearch.git
cd CoAutoResearch
node bin/auto-research.js doctor
node bin/auto-research.js uiOpen the printed URL and keep the server running. The dashboard needs no application dependencies or build step. Use --projects-dir /path/to/projects to keep research outside the checkout.
PDF generation additionally needs Python 3.12+, four pinned scientific skills, LaTeX and Poppler. Follow the agent setup guide. You can use this source checkout without an npm release.
Your first session: Create project → send a research brief → review the direction → Start autoresearch. Include your materials, compute budget and decisions that require approval. Creating a project alone does not start research.
Each Trial is a bounded research iteration. The service checks proposed changes, records accepted results and determines whether to continue, pause or request a human decision. Internal review is not external peer review or proof of a scientific claim.
Add to research draft prepares text; it does not send or apply it. Pause after current turn requests a pause at an agent-turn boundary, possibly inside a Trial. Resume autoresearch continues from the retained state.
Illustrated workflow; timing is schematic. Static alternatives are available. SVG
 Can covariance shrinkage help classify handwritten digits with little training data? Matched small-sample comparisons and a single held-out evaluation; findings remain specific to this dataset and protocol. |
 How reliably can a laptop estimate an Ising phase transition? CPU-bounded estimation and implementation diagnosis. Partial gains did not meet the joint criteria; held-out seeds stayed unused. |
Both papers are developer-operated research drafts with human intervention, not external user case studies. Their limitations remain in the reports; human scientific review is required before publication.
When reviewed results are recorded and project agents are idle, choose Generate paper. The internal agent uses writing, visualization, citation and venue skills on a frozen evidence snapshot, without running new experiments.
Open Paper to zoom, scroll and download the PDF and source. A previous draft remains available while its replacement is generated.
Example papers · Paper generation
The linked guides are in English. Start with Setup or Walkthrough; consult FAQ for common questions and Platforms for support boundaries.
Help verify installation on your platform, report reproducible bugs, improve research workflows, maintain translations, or share documented research outcomes. Use Issues and read the contribution guide; report security issues privately.
Issues · Contributing · Translations · Security · Code of conduct
Cite CoAutoResearch
Cite the software and record the version or commit used.
@software{coautoresearch,
title = {CoAutoResearch},
author = {CoAutoResearch contributors},
url = {https://github.com/YihongT/CoAutoResearch}
}Licensed under Apache 2.0. Upstream tools and skills retain their own licenses and attribution requirements.
Contact: yihong.tang@mail.mcgill.ca