OpenScience is an open-source AI workbench automating the scientific research loop for ML, biology, physics, and chemistry researchers. Users define a goal, and the AI handles literature review, hypothesis generation, code execution, experimentation, database querying, and report writing. It provides a browser-based workspace and supports diverse AI models via local API keys, acting as a collaborative research assistant.

How It Works

A local server hosts the browser UI, agent runtime, and tools. Agents (generalist and specialist) plan tasks using a harness, leveraging over 290 skills for training, evaluation, scientific domains, and cloud compute. It integrates direct querying for 30+ scientific databases (UniProt, PDB, arXiv). Its model-agnostic design allows seamless switching between providers or local models using user-provided API keys, which are managed locally.

Quick Start & Requirements

Install via npm (npm install -g @synsci/openscience) or npx (npx synsci). Run openscience to launch the browser workspace. Development requires Bun 1.3+. Users must provide AI model API keys (e.g., export ANTHROPIC_API_KEY=sk-ant-...). Docs: openscience.sh/docs.

Highlighted Details

• Skills: 290+ skills covering training (DeepSpeed, PEFT), evaluation, biology, cheminformatics, LaTeX, figures, and cloud compute (Modal, Tinker).
• Databases: Direct query access to 30+ scientific databases (UniProt, PDB, ChEMBL, PubChem, arXiv).
• Workspace: Browser UI with file tree, editor, terminal, history, and inline rendering for scientific data.
• Model Agnosticism: Supports any frontier/open-weight model via local API keys; no vendor lock-in for BYOK.
• Extensibility: LSP integration, custom agents, plugins, TypeScript SDK.

Maintenance & Community

Contribution guidelines, architecture, agent style guides, and code of conduct are documented. Specifics on active contributors, sponsorships, or community channels (Discord/Slack) are not detailed in the README.

Licensing & Compatibility

Licensed under Apache License 2.0. Compatible with various AI models and scientific databases. Bring-your-own-key usage is free. The project is independent and not affiliated with Anthropic.

Limitations & Caveats

The AI agent is not sandboxed; the permission system is for awareness, not isolation. Users needing strict isolation should use containers or VMs. Credentials are filtered and redacted from output.