An AI co-scientist that reads papers, codes, and runs experiments
It’s a browser-based AI workbench that tries to automate the full research loop — literature, code, experiments, and write-ups — using whatever models and compute you already have.
What it does
OpenScience is a local AI workbench that runs in your browser and attempts to automate end-to-end scientific research. You give it a goal, and it plans a project, searches literature through tools like arXiv and Semantic Scholar, queries biological and chemical databases, writes and executes code, and drafts results. It is model-agnostic, routing requests per-call to Anthropic, OpenAI, Google, or local models using your own keys, with no required account.
The interesting bit
The project’s real weight is in its integration layer rather than chat novelty. It ships with 290+ skills — from DeepSpeed and PEFT to cheminformatics and clinical biology — and native connectors to around 30 scientific databases including UniProt, PDB, and ChEMBL. That turns the LLM from a text generator into a collaborator that can actually look up protein structures or spin up training jobs.
Key highlights
- Full-loop research agent: literature review, hypothesis, code, experiment, analysis, and write-up in one session.
- 290+ built-in skills covering ML training stacks, molecular biology, dataset work, LaTeX, and cloud compute.
- Native scientific database tools: UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, Semantic Scholar, and more.
- Browser workspace with file tree, terminal, session history, and inline rendering for molecules, genomes, and plots.
- Extensible via MCP servers, LSP, custom agents, and a TypeScript SDK; model routing is per-request, so swapping providers mid-session requires no reconfiguration.
Caveats
- The agent is explicitly not sandboxed; the permission system is for visibility, not isolation. The README advises running inside a container or VM if you need security boundaries.
Verdict
Worth a look if you’re a researcher in ML, biology, physics, or chemistry who wants an autonomous coding partner with real database access and local compute. Skip it if you’re after a simple chat wrapper or are unwilling to let an unsandboxed agent execute commands on your machine.
Frequently asked
- What is synthetic-sciences/openscience?
- It’s a browser-based AI workbench that tries to automate the full research loop — literature, code, experiments, and write-ups — using whatever models and compute you already have.
- Is openscience open source?
- Yes — synthetic-sciences/openscience is open source, released under the Apache-2.0 license.
- What language is openscience written in?
- synthetic-sciences/openscience is primarily written in TypeScript.
- How popular is openscience?
- synthetic-sciences/openscience has 796 stars on GitHub.
- Where can I find openscience?
- synthetic-sciences/openscience is on GitHub at https://github.com/synthetic-sciences/openscience.