synthetic-sciences/openscience

OpenScience – An AI‑powered research workbench

What it is

• A desktop, web‑based, or command‑line application that lets you describe a scientific task in plain English. The built‑in research agent then plans the work, searches literature, writes and runs code (Python, R, shell), executes experiments on local or remote compute, and finally produces a report. Every step is logged so you can inspect exactly what the agent did.

Key capabilities

Capability	How it works
Literature review	Searches databases such as PubMed, arXiv, ChEMBL, UniProt and returns cited evidence.
Data analysis	Reads your files, generates cleaning/visualisation scripts, runs them in Python/R kernels, and saves figures and reproducibility code.
Reproduce experiments	Takes a claim, gathers prerequisites, runs the same analysis on your hardware (or on Slurm/SSH clusters) and compares results.
Compute orchestration	Uses local kernels for quick jobs; can dispatch to Modal, SSH hosts, or HPC schedulers (Slurm, PBS).
Skill library	Over 200 pre‑bundled “skills” – procedural knowledge for biology, chemistry, physics, ML, data engineering, etc. Users can add custom skills or plugins via a TypeScript SDK.
Model flexibility	Bring your own LLM API key, connect to a local model (Ollama, LM Studio), or use the managed “Ace” service.
Permission model	Granular grants for file reads/writes and network calls; risky actions require explicit approval.

Installation & usage

• Desktop app – download installers for macOS, Windows, Linux (auto‑updates).
• CLI – npm i -g @synsci/openscience then openscience. You can also run via npx synsci or a curl‑pipe installer.
• First‑run example:

  openscience ~/research/my-project
  # then type a natural‑language instruction, e.g.
  Inspect data/samples.csv for missing values, save a quality report and plot in results/.
  

• Commands like openscience run "Review the analysis plan" let you script single‑turn actions.

Extensibility

• Plugins & custom agents via the TypeScript SDK (generated from an OpenAPI contract).
• NVIDIA BioNeMo adapters – ten pre‑configured endpoints (e.g., DiffDock, OpenFold) for protein‑design tasks, requiring your own NVIDIA API key.
• Skill creation – add new procedures to the skill directory; the repo includes docs on how to contribute.

Benchmarks

• Evaluated on Terminal‑Bench Science (75.7 % success), BiomniBench‑DA (82.2 score) and a science subset of Terminal‑Bench 4.0 (10/14). Benchmarks and trace logs are in a separate benchmarks-openscience repo.

Documentation & community

• Full docs at openscience.sh/docs covering quick‑start, workflows, permissions, automation (JSON output, SDK), and troubleshooting.
• Open source under Apache 2.0; contributions accepted via GitHub issues and pull requests.

Bottom line OpenScience is a genuine, open‑source AI research assistant that combines LLM‑driven planning with real scientific tools (shell, Python/R, HPC, database connectors). It aims to automate the tedious, reproducible parts of scientific work while keeping the researcher in control of interpretation and judgment.

---

All information above is taken directly from the repository’s README.