synthetic-sciences/openscience
The open-source AI workbench for scientific research
What it solves
OpenScience is an AI-powered workbench designed to automate the scientific research loop. It eliminates the manual effort of switching between literature reviews, coding, running experiments, and writing results, by providing a unified environment where an AI agent can handle these tasks autonomously.
How it works
The system runs a local server that hosts a browser-based workspace UI, an agent runtime, and a tool layer. The agent uses a research harness to plan and execute tasks by calling tools such as a shell, editor, LSP, and specialized scientific connectors. It is model-agnostic, allowing users to provide their own API keys for models from providers like Anthropic, OpenAI, and Google.
Who it’s for
It is built for researchers in machine learning, biology, physics, and chemistry who want an AI collaborator to help them conduct research, query scientific databases, and manage experiments.
Highlights
- End-to-end research loop: Handles literature review, hypothesis formation, code execution, and final write-ups in a single session.
- Specialized agents: Includes a general research agent and specialists for biology, physics, and ML, supported by critique and literature-review sub-agents.
- Extensive toolset: Features over 290 skills across training (DeepSpeed, PEFT), cheminformatics, and cloud compute (Modal, Tinker).
- Scientific database integration: Direct agent access to over 30 databases, including UniProt, PDB, arXiv, and PubMed.
- Integrated workspace: A browser UI with a file tree, terminal, and inline rendering for molecules, genomes, and plots.