Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Open Science Desktop

A local-first desktop research workbench: bring any model, and agents run surveys, experiments and papers with every file traced back.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.9/5

Open Science Desktop is a serious open attempt at a desktop AI lab notebook, with real provenance and sensible approval defaults, but it is beta software doing powerful things on your own machine.

What is Open Science Desktop?

Open Science Desktop is an open-source research workbench for macOS, Windows and Linux, built with Tauri 2, React and TypeScript. It wraps a bundled, pinned OpenCode runtime (OpenCode 2 since v0.7.2), so you choose the model: any provider OpenCode supports, an OpenAI-compatible endpoint or a local one. Claude Code, Codex and other Agent Client Protocol agents can also run inside it.

Its research skills cover the whole loop. research-explorer turns a broad direction into topics, literature-survey writes a 6 to 20 page PDF with 60 or more citations, experiment-suite builds runnable code with results and figures, and paper-writer drafts an 8 to 14 page paper; ai4s-agent chains them, and integrity-auditor checks a paper's images, numbers and logic. Every step lands as files in your workspace, and .openscience/provenance.jsonl links each artefact to the run or edit that made it.

Around that sit one-click MCP connectors (arXiv, PubMed, Crossref, Semantic Scholar, bioRxiv, ClinicalTrials.gov, Materials Project, FRED and others), Jupyter notebooks, runs on local machines, SSH and Slurm hosts or Modal, control of your own Chrome, computer use for desktop apps such as ImageJ, a headless server (osd) and a token-protected gateway for phone access. It is MIT-licensed; v0.7.3 shipped on 16 November 2026.

Is Open Science Desktop worth it?

Open Science Desktop is aimed at researchers who want an agent to do the drudge work of a project while leaving an audit trail a reviewer could follow. Its strongest idea is that nothing lives only in the chat: surveys, code, results.json files, figures and reports are files in a local folder, and the provenance log ties each one to the code, inputs and conversation that produced it. It has about 2,113 GitHub stars, and the project says it ranked first by scored-task average on ResearchClawBench in July 2026, a claim from its own README.

The skill outputs are concrete and checkable: a literature survey with 60-plus citations and LaTeX source, an experiment package with a design document and runnable code, a citation review that resolves DOIs and flags unsourced numbers. Being model-agnostic through OpenCode means a lab can keep data on a local model or switch providers without changing tools.

The defaults are careful for software this capable. The desktop app asks before running commands, deleting files, installing dependencies or reaching the network; turning that off (osd approval set full) has to be done on the machine itself. The remote gateway is loopback-only until you opt in to LAN access, and API keys stay in app-private config, outside the workspace and exports. Browser control can drive your real Chrome profile with its logins, and computer use can operate any app on screen, so those deserve the same care as giving a colleague your laptop.

The rough edges are stated plainly. The README calls it beta research tooling whose output should be treated as drafts. Windows and Linux builds are not code-signed yet (macOS builds are signed and notarised), the headless server does not write provenance.jsonl or the run index, and Portuguese and Arabic interfaces are registered but not selectable. Releases have landed every nine or ten days lately (v0.7.1 on 29 October, v0.7.2 on 7 November, v0.7.3 on 16 November 2026), so expect change.

Where it earns its place

  • A PhD student in computational biology

    Get a literature survey on protein language models to start a thesis chapter

    Run literature-survey with the PubMed, bioRxiv and Semantic Scholar connectors; it compiles a 6 to 20 page PDF with LaTeX source, and the citation review checks that DOIs resolve. The app is free; model costs depend on the provider chosen.

  • A machine learning lab with a shared Slurm cluster

    Rerun a benchmark sweep and keep a record reviewers can check

    Register the cluster from ~/.ssh/config, let experiment-suite write the code and submit jobs, and keep the run records and provenance log with the paper's repository. Approvals stay on for commands and network access.

  • A materials scientist on a lab workstation

    Pull data from instrument software into figures without manual export

    Use computer use to drive the desktop instrument app through its accessibility tree, then generate figures in a notebook; each action reports whether it could be verified. On Windows, expect a SmartScreen warning because the build is unsigned.

Before you commit

Setup
Download the installer from GitHub Releases (macOS 13+, Windows 10/11 x64, .deb or .rpm on Linux), add a model provider key or a local endpoint, and start a session; servers can use the self-contained osd archive.
Integrations
OpenCode provider catalogue and OpenAI-compatible endpoints; MCP connectors for arXiv, PubMed, Crossref, Semantic Scholar, bioRxiv/medRxiv, ClinicalTrials.gov, MyVariant/ClinVar, Materials Project, FRED and Open-Meteo, plus any MCP server; ACP agents such as Claude Code, Codex and Gemini CLI; Zed, JetBrains and Neovim via ACP; Jupyter, SSH/Slurm and Modal.
Data and privacy
Workspace files, sessions, provenance and run records stay local by default; prompts and files go to whichever model provider you pick, and Settings shows what can be sent. Provider keys are kept in app-private config, not in the workspace or exports.
Cost at scale
No licence cost; the bill is model tokens, which long survey and paper-writing runs use heavily, and remote compute. The project publishes no token figures.
Limits
Beta MVP; unsigned Windows and Linux builds; no provenance from the headless server; local Jupyter kernels and native dialogs absent without a desktop; a folder's MCP configuration is cached on first use until the server restarts.

What’s good

  • Every artefact traced to its code, inputs and conversation
  • Any model through the bundled OpenCode runtime, including local ones
  • Survey, experiment and paper skills with concrete outputs
  • Approvals on by default; gateway loopback-only until enabled
  • SSH, Slurm and Modal runs recorded as reproducible run records

What’s not

  • Beta; outputs need checking before publication
  • Windows and Linux builds are unsigned
  • Browser control and computer use act with your own logins
  • Headless server skips the provenance log
  • Benchmark lead is self-reported

Best for

Research groups who want a local, model-agnostic agent workbench that keeps a provenance trail for surveys, experiments and drafts.

Open Science Desktop alternatives

Open Science Desktop: common questions

Is Open Science Desktop worth it?
Open Science Desktop is a serious open attempt at a desktop AI lab notebook, with real provenance and sensible approval defaults, but it is beta software doing powerful things on your own machine. We rate it 3.9 out of 5.
How much does Open Science Desktop cost?
Free and open source (MIT). You pay your model provider, or nothing with a local model, plus any SSH, Slurm or Modal compute you use.
What are the downsides of Open Science Desktop?
Beta; outputs need checking before publication; windows and Linux builds are unsigned; browser control and computer use act with your own logins; headless server skips the provenance log; benchmark lead is self-reported.
Who is Open Science Desktop best for?
Research groups who want a local, model-agnostic agent workbench that keeps a provenance trail for surveys, experiments and drafts.
How much does Open Science Desktop cost for a team or in production?
No licence cost; the bill is model tokens, which long survey and paper-writing runs use heavily, and remote compute. The project publishes no token figures.
Is Open Science Desktop safe for company data?
Workspace files, sessions, provenance and run records stay local by default; prompts and files go to whichever model provider you pick, and Settings shows what can be sent. Provider keys are kept in app-private config, not in the workspace or exports.
What are the best Open Science Desktop alternatives?
The closest alternatives we have reviewed are The AI Scientist (v1, v2 and v3), GPT Researcher, OpenResearch and OpenCode.

Answers checked Nov 20, 2026.