What if you could give an AI a research goal and have it return a complete write-up with real experimental results, all without human intervention between start and finish? OpenScience, an open-source project from Synthetic Sciences, does exactly that. With 2,569 stars on GitHub in just over two weeks since its July 3 launch, it is one of the fastest-growing research tools in the open-source ecosystem. It runs the full scientific research loop autonomously: reading relevant literature, forming testable hypotheses, writing and executing code, querying major scientific databases, and producing a written summary of findings. All from a single browser-based workspace powered by your own API keys.
What OpenScience Does
Most AI research tools focus on one part of the pipeline. Elicit helps with literature review. PaperQA answers questions about papers. Claude Science can write and run Python in a sandbox. OpenScience is different because it strings the entire loop together into one continuous session. You give it a goal, and it works through the steps a human researcher would: understanding the landscape, forming a hypothesis, designing an experiment, running code on real compute, analyzing results, and writing up what it found. It is not a chatbot that answers questions about science. It is an agent that does science.
The system runs as a local server that hosts the workspace UI, the agent runtime, and the tool layer. The agent plans with a research harness, calls tools for shell commands, file editing, LSP integration, MCP servers, scientific database connectors, and its built-in skill library. Models are routed per request, so you can switch between Anthropic, OpenAI, Google, or any open-weight provider between tasks without changing anything else. Sessions, artifacts, and provenance data are stored on disk and can be shared as links.
290 Built-in Skills and 30 Scientific Databases
OpenScience ships with more than 290 pre-built skills spanning machine learning training (DeepSpeed, PEFT, TRL), evaluation, dataset preparation, molecular biology, clinical biology, cheminformatics, papers and LaTeX, figure generation, and cloud compute integration. These are not thin wrappers around APIs. They are purpose-built research capabilities that the agent calls as tools during its workflow. The skill library is one of the project's strongest differentiators because it means the agent can actually do real work in domains like protein structure analysis, drug discovery, and materials science without needing a human to configure each step.
The scientific database integration is equally extensive. OpenScience connects to UniProt, the Protein Data Bank, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, Semantic Scholar, and roughly 30 more databases. The agent queries these directly as part of its research loop, pulling real data and incorporating it into its analysis. For a biologist studying a protein family, for example, the agent can pull sequences from UniProt, structures from the PDB, and relevant literature from Semantic Scholar in a single session. The model-agnostic design means the agent can choose the cheapest frontier model for each subtask, keeping costs manageable for independent researchers and small teams.
How It Compares to Alternatives
The existing landscape of AI research tools is fragmented. Elicit and scite specialize in literature review but do not run experiments. Google's AlphaFold and AlphaProteo solve specific structural biology problems but do not generalize across domains. Anthropic's Claude Science can write and execute code in a sandboxed environment but lacks the skill library and database connectors for specialized scientific work. OpenScience's advantage is breadth: a single platform that covers the entire loop across biology, physics, chemistry, and machine learning, all under an Apache 2.0 license with no vendor lock-in.
That breadth comes with trade-offs. The agent is not sandboxed, and the permission system is designed for awareness, not isolation. The README explicitly recommends running inside a container or VM for sensitive workloads. The 290 skills are powerful but unevenly documented; some are clearly battle-tested while others are still maturing. And setting up the full workflow with all database connectors requires API keys and configuration that a non-technical researcher might find intimidating. For a solo founder or small team with technical depth, however, the trade-off is clearly worth it.
The project's architecture follows the same pattern that made VS Code successful: a local-first core with extensibility through plugins, MCP servers, and a TypeScript SDK. The monorepo structure is clean and well-organized, with the CLI and server in backend/cli, the workspace UI in frontend/workspace, and documentation in frontend/docs. Development requires Bun 1.3 or newer, and the build pipeline produces platform binaries for macOS, Linux, and Windows.
Who This Is For
OpenScience is built for researchers and teams who want AI-powered research automation without handing their data or workflows to a third-party vendor. Three groups will benefit most. First, academic research groups with limited compute budgets: the bring-your-own-key model means you pay only for the API calls you make, and the model-agnostic design lets you use cheaper open-weight models for routine tasks. Second, biotech and pharma startups looking to accelerate early-stage discovery: the skill library and database connectors reduce the time from literature review to experimental design from weeks to hours. Third, solo technical founders building in science-adjacent spaces: OpenScience eliminates the need to hire dedicated ML engineers for research automation. One researcher with an API key and a browser can run what previously required a team of five to ten people. With 2,569 stars and growing fast, this is a project worth watching for anyone building at the intersection of AI and scientific research.