DeepScientist is a local-first AI research studio, Bring your own AI scientist onto your machine in 15 minutes.
GitHub | Chinese README | English Docs | Paper | Website
15-minute local setup · One repo per quest · Visible research progress · Human takeover anytime
Quick Start • Launch Your First Project • Product Tour • Model Setup
If you are tired of paper overload, broken baselines, scattered experiment logs, and late-night writing cleanup, give the project a star first, then keep scrolling to see how much research grunt work it can take off your plate.
deepscientist.mp4
What drains researchers is often not the lack of ideas. It is the endless cycle of low-leverage work:
- new papers keep coming, but only a small fraction turns into an actionable next-step research plan
- baseline repos fail on environment, dependency, data, and script issues before real work even starts
- experiment results get scattered across terminals, scripts, notes, and chats, making later review painful
- writing, figures, and analysis live in separate tools, so turning them into a coherent paper takes far too long
This is the problem DeepScientist is built to solve:
turn fragmented, repetitive, easy-to-lose research work into a local AI workspace that can keep moving, keep accumulating, and keep getting stronger over time
It is not a tool that summarizes papers, throws you a few ideas, and leaves the dirty work to you.
It is much closer to a real long-running AI research partner:
| What common AI tools often look like | What DeepScientist does instead |
|---|---|
| Great at chatting, but context disappears quickly | Turns tasks, files, branches, artifacts, and memory into durable state |
| Good at suggesting ideas, but weak at sustained execution | Pushes papers, baselines, experiments, and writing inside one workspace |
| Strong automation, but feels like a black box | Lets you inspect the process through the web workspace, Canvas, files, and terminal |
| Hard to take over once it goes off track | Lets you pause, take over, edit plans, change code, and continue at any time |
| Each run ends when the run ends | Preserves failed paths, winning paths, and reproduction lessons for the next round |
DeepScientist is not a one-shot agent demo. It is a system built for long-horizon research work.
- feed it a core paper, a GitHub repository, or a natural-language research objective
- it turns those inputs into an executable quest instead of a chat that loses state after a few turns
- restore repositories, prepare environments, handle dependencies, and track the critical failures
- preserve what broke, what got fixed, and which steps are trustworthy for future rounds
- propose the next hypothesis from existing results
- branch, ablate, compare, and record conclusions
- keep failed routes as assets instead of deleting them
- organize findings, conclusions, and analysis
- produce figures, reports, and paper drafts
- support local PDF and LaTeX compilation workflows
- the web workspace in your browser
- the TUI workflow on a remote server
- external connector surfaces for collaboration and progress updates
The current docs already cover these collaboration channels:
What retains users is not a flashy demo. It is a system that becomes more useful the longer you work with it.
DeepScientist tends to stick for four reasons:
- code, experiments, drafts, and project state stay on your own machine or server by default
- this is especially valuable for unpublished ideas, sensitive experiment history, and longer-running research loops
- every quest is a real Git repository
- branches, worktrees, files, and artifacts naturally express research structure
- it does not only give you an output
- you can inspect what it read, what it changed, what it kept, and what it plans to do next
- DeepScientist can move autonomously
- you can also step in, edit, redirect, and hand control back whenever you want
Because this is not just a concept. It is a real system with public docs, a public paper, and a public install path.
2026/03/24: DeepScientist officially releasedv1.52026/02/01: the paper went live on OpenReview forICLR 2026- npm install path is already available:
@researai/deepscientist - both Chinese and English docs are available, along with Web, TUI, and connector entry points
[TODO asset] If you can later add one public case summary such as "reproduce a baseline + finish multiple experimental rounds + draft a paper package," this section will become much more convincing.
If you want to try it right now, the shortest path is:
npm install -g @researai/deepscientist
codex --login
ds --hereIf codex --login is unavailable, run this once first:
codexAfter startup, the default local address is:
http://127.0.0.1:20999
Then you only need to do three things:
- click
Start Research - fill in the research goal, baseline links, paper links, or local paths
- let DeepScientist start a real research project that can keep evolving locally
If this is your first run, prefer an isolated environment, a non-root user, and a local machine. For the full details, see:
- 15 Codex Provider Setup
- Weixin Connector Guide
- QQ Connector Guide
- Telegram Connector Guide
- WhatsApp Connector Guide
- Feishu Connector Guide
- graduate students and engineers who want to reproduce papers and push beyond existing baselines
- labs or research teams running long experiment loops, ablations, and structured result analysis
- people who want code, experiments, notes, and writing to live in one workspace
- users who do not want to hand unpublished ideas and intermediate results directly to a pure cloud workflow
- people who want to run work on servers while following progress from web, TUI, or messaging surfaces
We believe a system that is actually suitable for research should at least satisfy these principles:
- one quest, one repository, instead of letting everything dissolve after a short conversation
- branches and worktrees should express research routes naturally instead of being forced into chat history
- failed paths should be preserved, summarized, and reused instead of overwritten
- human researchers should always retain takeover power instead of being locked outside the loop
- the research process should be reviewable, inspectable, and auditable instead of relying on "the model says it did it"
If that sounds like the way you want to work, DeepScientist is worth trying now.
If you like DeepScientist, you may also want to explore the rest of the ResearAI ecosystem:
| Project | What it does |
|---|---|
| AutoFigure | generate publication-ready figures |
| AutoFigure-Edit | generate editable vector paper figures |
| DeepReviewer-v2 | review papers and suggest revisions |
| Awesome-AI-Scientist | curated AI scientist landscape |
If you are developing or maintaining DeepScientist, continue with:
If DeepScientist helps your research or engineering work, please cite:
@inproceedings{
weng2026deepscientist,
title={DeepScientist: Advancing Frontier-Pushing Scientific Findings Progressively},
author={Yixuan Weng and Minjun Zhu and Qiujie Xie and QiYao Sun and Zhen Lin and Sifan Liu and Yue Zhang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=cZFgsLq8Gs}
}If this feels like the research workflow you have been waiting for, give the project a star. Every star makes it easier for more researchers who actually need it to find it.