ai4s-research/open-science

Open Science Desktop – Quick Overview

What it is – A local‑first, model‑agnostic desktop workbench that lets you run the whole scientific‑research loop (idea → literature survey → experiment → analysis → paper) with AI agents. It works on macOS, Windows and Linux and can also be run head‑less via a CLI (osd).

Key capabilities

• Autonomous research agents – pre‑built skill chains (research‑explorer, literature‑survey, experiment‑suite, paper‑writer, etc.) that produce real artefacts (code, data, figures, PDFs) instead of just chat text.
• Full provenance – every file, figure or table is linked to the exact code, model output and conversation that created it, stored in a .openscience/provenance.jsonl log.
• Local‑first & privacy‑first – all sessions, data and keys stay on your machine unless you explicitly share them. API keys never leave the host.
• Model‑agnostic runtime – uses the OpenCode side‑car; you can plug in any OpenAI‑compatible endpoint or other providers via the built‑in SDK.
• Multi‑pane UI – split‑screen tiling, independent “Screens”, per‑pane model selection, and commands like /plan and /goal to structure work.
• Remote & mobile access – a token‑authenticated gateway serves the same UI to a browser on the LAN, phone, or via a tunnel; a head‑less osd server runs the workbench on headless machines.
• Browser & computer control – the agent can drive your own Chrome (preserving logins) or a private browser, and can interact with desktop apps through accessibility APIs.
• Extensible skills & connectors – skills are fetched from the ai4s-skills pack; one‑click connectors to literature databases (arXiv, PubMed, Crossref, etc.) and domain data sources (Materials Project, FRED, USGS, etc.).
• Notebook support – real Jupyter notebooks (.ipynb) with local kernel execution via bundled uv.
• Safety & reproducibility – run records are stored in SQLite, runs can be reproduced on remote clusters (SSH/Slurm, Modal) and include environment snapshots.

Installation – Download platform‑specific installers from the Releases page (macOS dmg, Windows exe/msi, Linux deb/rpm). The installer also places the osd CLI on your PATH.

Typical CLI workflow

# configure model & key (keys stay in memory)
osd auth set anthropic --key sk-…
osd model set anthropic/claude-opus-4-5

# start headless server (LAN token printed)
osd server --lan

# create a project & session, send a prompt, wait for completion
osd project new "Reef survey"
sid=$(osd session new --project "Reef survey")
osd session send $sid "Fit bleaching trend 2015‑2024 and write report.md" \
    --model anthropic/claude-sonnet-4-5 --wait

# fetch results
osd fs ls figures/
osd fs get report.md --output ./report.md

Status – Actively developed (v0.6.0 as of Oct 2026). Ranked #1 on the ResearchClawBench benchmark for autonomous scientific agents.

License & citation – MIT license; DOI 10.5281/zenodo.21351225 for citation.

---

All details are taken directly from the repository’s README.