Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM
Local AI Agents & Tool Use

Open Science Desktop Review: Free AI Research Workbench for Your Own Files

·8 min readBy · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

Open Science Desktop is a free, MIT-licensed desktop app for macOS, Windows and Linux that runs an AI agent over your own project folders to search literature, run code and notebooks, and produce figures and reports. It does not include a model: you add a provider key or endpoint, so what leaves your machine depends on that choice. The maker calls it beta research tooling.

Key Takeaways

  • Installers are published for macOS, Windows and Linux; the latest release is v0.7.3.
  • The app is free under the MIT license; you pay your model provider for usage.
  • The agent asks for approval before running commands, deleting files, installing dependencies or using the network.
  • The maker describes it as beta research tooling and says outputs must be verified before use.
  • PromptQuorum has not tested the app, so every capability below is a maker statement.

📍 In One Sentence

Open Science Desktop is a free desktop app in which an AI agent runs literature searches, code and notebooks on your local project folders and traces every figure and report back to its source.

💬 In Plain Terms

Think of it as a research workspace on your computer with an agent built in. Your files and history stay in local folders, but the prompts go to whichever model provider you connect.

📌Note: Research outputs from an AI agent can contain wrong numbers and invented citations. Check results yourself before you publish or decide anything. This is general information, not professional advice.

Open Science Desktop (formerly Open Science) is a free, MIT-licensed desktop app for macOS, Windows and Linux. It bundles an AI agent workbench for research: literature searches, code, notebooks, figures and reports, with project files and run history kept in local folders. You connect your own model provider, so the model itself usually runs in the cloud unless you point the app at a local endpoint. This review is based on the maker's website, the GitHub repository, its README and the release pages. PromptQuorum has not installed or tested the app.

Open Science Desktop Review: Free AI Research Workbench for Your Own Files

Overview

Open Science Desktop turns a research request into a plan, local execution and a set of linked outputs inside one desktop app.

The repository belongs to the ai4s-research organisation, which describes itself as sponsored and primarily maintained by APEVON TECHNOLOGIES PTE. LTD. (Singapore). The project was created on GitHub in July 2026, has more than 2,000 stars and was previously called Open Science. Version 0.7.3 was released on 16 November 2026 and the repository had a commit on 17 November 2026, so development is ongoing and no stale-project warning applies.

The app is built with Tauri 2 and React and starts a bundled OpenCode runtime in the background, separate from any OpenCode setup you already have. The README says it is a working desktop MVP in active development, and the maker calls it beta research tooling.

The maker describes it as an open-source desktop alternative to Claude Science and similar workbenches, and its README claims first place on the ResearchClawBench leaderboard. PromptQuorum has not compared it with other tools and has not checked that ranking.

A different project called OpenScience, from Synthetic Sciences, is an independent Apache-2.0 command-line tool that opens a browser workspace. The maker's own comparison page says the two are separate projects. This review covers only Open Science Desktop.

How to Get It

You install Open Science Desktop with a normal installer from GitHub Releases, which the maker's download page also links.

The v0.7.3 release lists .dmg files for Apple Silicon and Intel Macs, a Windows .exe setup file plus an .msi for IT-managed deployment, and .deb and .rpm packages for x86_64 and ARM64 Linux. The README says the macOS packages are signed and notarized, while the Windows and Linux builds are not signed yet. Pick one Windows format and stay on it.

Platform
How to get it
macOS.dmg (Apple Silicon or Intel), macOS 13+
Windowsx64 setup .exe (or .msi for IT); SmartScreen warning
Linux.deb or .rpm, x86_64 and ARM64
Linux, no rootInstall script from the README (terminal)
WebsiteOfficial site with download page and docs
Privacy policyNo policy page found; see the Privacy section

Normal users should use the installers. Building from source needs Node.js, pnpm, Rust and the Tauri prerequisites. The app needs a model provider key or endpoint before the agent can answer.

Features

The core is an agent that works on files in a local workspace and records where every output came from. All items below are maker statements from the README, website and release notes.

  • Research skills that chain exploration, literature survey, experiments and paper writing, plus an integrity-audit skill.
  • Provenance: figures, tables and reports link to the code, inputs and run that produced them, with append-only run logs.
  • Real Jupyter notebooks for Python and R with a local kernel, and viewers for PDF, CSV, DOCX, XLSX and other formats.
  • Runs on remote machines through SSH or Slurm, or on Modal.
  • One-click connectors for arXiv, PubMed, Crossref, Semantic Scholar and bioRxiv, plus any MCP server you add.
  • Claude Code, Codex and other agents can run beside the bundled runtime, and the app speaks the Agent Client Protocol in both directions.
  • Browser control through your own Chrome profile, and computer use of desktop apps through accessibility APIs.
  • An opt-in gateway that serves the same interface to a browser or phone, and an osd command for headless servers and scripts.
  • Interface languages: English, Simplified Chinese, Japanese, Spanish, German, French and Korean.

Privacy and Data Flow

Workspace files, session history and provenance stay in local folders by default, but prompts and any content the agent reads go to the model provider you select.

The README says command execution, file deletion, dependency installation and remote connections are approved by the user, and that provider credentials are stored in app-private runtime configuration. It also says Settings contains a plain-language data-flow view of what can be sent to the selected provider. The osd command has a setting called full approval that turns these prompts off, and the maker says it is a deliberate choice, not a default.

Two features widen what the agent can reach. Browser control uses your real Chrome profile with its logins, and computer use can read and operate apps on your screen. The remote-access gateway is off by default and listens on the local machine until you opt in.

No privacy policy, terms or imprint page was found on the maker's website. The sponsor page names a contact address for privacy questions, and whether the app sends telemetry or analytics is not stated in the sources read.

Who Should Use It

Open Science Desktop fits people who want an agent workbench on their own computer and can manage a provider key.

  • Researchers and students who want literature, code, figures and reports kept together with traceable sources.
  • Developers who want to run several coding agents and a research agent in one window.
  • Teams that need a headless server or a browser-accessible workbench and will configure access themselves.

What We Could Not Verify

Open Science Desktop is not a fit for everyone, and several details remain unverified.

  • Not a fit: anyone who needs a fully offline tool without connecting any model provider.
  • Not a fit: users who need signed Windows or Linux installers today; the README lists Windows code signing as near-term work.
  • Unverified: whether a local model works well through the app. The README mentions custom OpenAI-compatible endpoints, but PromptQuorum did not run one.
  • Unverified: the ResearchClawBench ranking and the quality of generated papers, surveys and citations.
  • Unverified: hardware requirements beyond the operating system, account requirements and telemetry.
  • Unverified: a privacy policy. None was found, so the privacy statements here come from the README.
  • Unverified: the maker's website lists v0.7.2 and an unsigned-macOS workaround, while GitHub shows v0.7.3 and signed macOS builds. This review follows GitHub.
  • Unverified: the repository license shows as Other on GitHub; the MIT text comes from the LICENSE file, and bundled third-party skills keep their own licenses.

Open Science Desktop vs Alternatives

Open Science Desktop is a desktop research workbench with a bundled agent runtime. PromptQuorum has not compared it hands-on with the tools below. The descriptions come from the directory entries and the maker's own comparison page.

Open Science Desktop

Noted in sources:
MIT desktop workbench; bring your own model

Vane

Noted in sources:
Self-hosted answering engine with cited answers

OpenScience (Synthetic Sciences)

Noted in sources:
Independent Apache-2.0 CLI opening a browser workspace

Compare in directory

FAQ

Is Open Science Desktop free?

The app is free under the MIT license. The download page says provider usage may incur charges, so the cost of the model you connect is separate and set by that provider.

Does it run models on my computer?

Not by itself. It bundles an agent runtime, not a model, and you add a provider key or endpoint. The README mentions custom OpenAI-compatible endpoints, but PromptQuorum did not test a local model.

Is it the same as OpenScience from Synthetic Sciences?

No. That is an independent Apache-2.0 command-line project that opens a browser workspace, according to the maker's comparison page. Open Science Desktop was previously named Open Science, which is why the names look alike.

Can I use it without a screen or from my phone?

Per the README, the osd command runs the workbench on a machine with no display, and an opt-in gateway serves the same interface to a browser or phone. It is off by default and local-only until you enable it.

How often is it updated?

The releases page shows v0.7.0 on 26 October, v0.7.1 on 29 October, v0.7.2 on 7 November and v0.7.3 on 16 November 2026, so four releases shipped within about three weeks.

Verdict

Open Science Desktop is worth a look if you want a free, inspectable research workbench on your own computer and already use a model provider.

On paper it covers a lot: local project folders, provenance links, notebooks, remote runs, connectors, approval prompts and MIT-licensed code. Releases arrive often, which suggests an active project.

The trade-offs are real. It is beta software, the model comes from a provider you choose, the Windows and Linux installers are unsigned, and no privacy policy was found. PromptQuorum has not tested it. Try it on a non-sensitive project, keep the approval prompts on, and check the data-flow view in Settings before you rely on it.

Sources

← Back to Power Local LLM

MCP