Open Science Workbench
Local-first, model-agnostic AI research workbench for reproducible scientific discovery.
Evidence shows permission controls (tool-level allow/ask/block), user confirmation (approval modes), data flow transparency (activity history, provenance views), sensitive data handling (encrypted keys, dropped env vars). But dependency security only mentions reporting, no specific audit; rollback only mentions branching and versions, no explicit rollback mechanism; source attribution only mentions author, no verification.
Self-consistency is good, README and code align; dependency availability not verified, only listed; failure messages not detailed.
Audience clear (researchers), scenarios broad; capability boundaries stated (sandbox not implemented); trigger precision moderate, no explicit trigger words; environment fit good (multi-platform).
Information architecture clear, install notes detailed, naming stable, examples and FAQ present, known limitations explicit, license Apache-2.0, versioning and changelog exist, maintenance responsibility not explicit.
Output usability high (reproducible reports, tables, figures), marginal value high (integrates many tools), cost-benefit not fully explained.
Claim traceability moderate, no specific evidence links; cross-source corroboration insufficient; fact-inference separation unclear.
- Sandboxing and network allowlisting not implemented; permission control relies on approval, handle sensitive data with caution.
- Dependency security lacks specific audit, supply chain risk needs attention.
- Rollback mechanism unclear; branching and versioning may not suffice for error recovery.
What does this agent do, and when should you use it?
Open Science is an open-source, local-first, model-agnostic AI research workbench that enables reproducible scientific discovery on your own computer (macOS, Windows, Linux). It integrates project management, multi-model agent execution, Python/R notebooks, scientific data connectors, immutable artifact versions with provenance, and permissioned human-in-the-loop controls into one workspace. Includes 18 featured research skills and 24 built-in scientific connectors, covering bioinformatics, computational biology, biomedical research, genomics, structural biology, and computational drug discovery. The app is distributed as an Electron desktop application with installers for macOS, Windows, and Linux. It uses an ACP-based agent runtime supporting Claude Code, OpenCode, or Codex, and offers optional localhost web UI and headless CLI/SDK modes.
Open Science lets you create a project, start a session, and describe a research task in natural language. The agent reads files, runs Python and R code, executes shell commands, searches the web, and calls scientific data connectors (e.g., PubMed, bioRxiv, ChEMBL) to generate reports, tables, and figures as artifacts. Each artifact is saved as an immutable, checksummed version with a Provenance view showing producer code, execution history, input references, environment inventory, and conversation branch. Users control the agent's actions via approval profiles (Ask for approval, Auto-approve edits, Full access). The interface supports @-references to files, /-selection of skills, and branching by editing past messages. Artifacts can be exported as .ipynb, and project files (up to 10GB) can be uploaded and previewed in formats including scientific data, PDFs, Office documents, images, source code, and molecular structures.
- A computational biologist uses built-in skills like AlphaFold2 and Boltz for structure prediction and manages protein files locally.
- A biomedical researcher conducts literature reviews using PubMed and Clinical Trials connectors, generating traceable review reports.
- A genomics researcher analyzes genetic variants with Genomes and Variants connectors and uses Evo 2 or scGPT for sequence analysis.
- A drug discovery scientist screens compounds via ChEMBL and ZINC connectors and designs molecules with DiffDock and LigandMPNN.
- A data scientist runs Python/R notebooks locally, manages analysis pipelines with the project library and artifact provenance.
- A researcher submits long-running jobs to a remote HPC cluster via the Remote Compute (SSH) skill and automatically analyzes results on completion.
What are this agent's strengths and limitations?
- Local-first architecture: projects and state are stored locally, enabling offline access and clear data ownership.
- Model-agnostic: supports multiple cloud providers, custom gateways, and Claude/Codex subscriptions, with easy model switching.
- Reproducibility: artifacts have immutable versions with provenance tracking, including producer code and execution history.
- Rich built-in skills and connectors for bioinformatics and drug discovery, with an extensible architecture.
- Requires an Electron desktop app installation, not a pure web service, which may be less suitable for server-only environments.
- Dependent on third-party model providers, requiring API keys or subscriptions and incurring usage costs.
- Advanced features like MCP connectors and remote HPC require configuration and skill enablement, adding a learning curve.
- Some features (e.g., deterministic reconstruction, portable environment restoration) are still on the roadmap and not fully implemented.
How do you install or deploy this agent?
Download the installer for your platform from the GitHub Releases page (macOS Apple Silicon/Intel, Windows x64, Linux x64 AppImage or Debian package). On first launch, complete the five-step setup: Environment checks, Agent runtime (select and optionally auto-install Claude Code, OpenCode, or Codex), Model provider (built-in cloud provider, custom gateway, or Claude/Codex subscription), Notebook runtime (optional, Python and R), and Data location. For development, clone the repo, run 'npm install' and 'npm run dev'.
How do you use this agent?
After installation, create a project and start a session, describe the research goal, input data, and expected outputs. Attach source files, select a verified model, and choose an approval mode, then send the task. In the session, use '@' to reference project files and '/' to select enabled skills. Create branches by editing previous messages, and inspect artifact provenance via the Provenance view. You can also install the CLI from Settings → General → Command line tool to run tasks headlessly, e.g., 'open-science start', 'open-science project create', 'open-science run --prompt-file task.md', and 'open-science artifacts download'.