Synthetic Sciences Releases OpenScience: An Open-Source, Model-Agnostic AI Workbench for Machine Learning, Biology, Physics, and Chemistry Research

Synthetic Sciences has been released OpenSciencean open source AI workbench for scientific research. Licensed under Apache 2.0 and it works on your infrastructure. The research team presents it as an open alternative to Anthropic’s Claude Sciencelaunched in late June 2026.

The tone is direct. AI science tools shouldn’t belong to a single vendor. OpenScience stores open workflows, model changes, and spatial data. It is an independent project, not affiliated with or endorsed by Anthropic.

The TL;DR

• OpenScience is an Apache-2.0, model-agnostic AI benchmark for machine learning, biology, physics, and chemistry.
• It uses a complete loop: literature, hypothesis, code, test, analysis, and writing.
• Any model works (Claude, GPT, Gemini, GLM, Kimi, DeepSeek, good local songs); change is per request.
• It ships with 250+ programmable capabilities, and databases (UniProt, PDB, ChEMBL, arXiv, and ~30 more) as agent tools.
• It works with your infrastructure with your keys; use of the bring-your-own key is free and never gated.

What is OpenScience

OpenScience is a browser-based workspace supported by a local agent runtime. He gives it a research goal. It then runs through a loop that a skilled programmer will follow.

Reads relevant papers, builds theory, writes and runs code, and runs tests. It queries the scientific database and writes the result. All of this happens in one continuous session.

The tool is model-agnostic by design. It works with any border or open weight model, using your own API keys. No account is required to get started.

Installation uses npm. The command says openscienceand opens a workspace in your browser.

npm install -g @synsci/openscience
openscience

The startup offers three options: Atlas-managed models, your provider keys, or free demo models. You can also skip the global installation. Running npx synsci he does the same in one step.

How It Works

OpenScience uses a local server. That server hosts the workspace UI, the agent runtime, and the tools layer. The agent arranges for a research harness and telephone tools.

Those tools include shell, editor, LSP, MCP servers, scientific connectors, and capabilities. The agent broadcasts its work back to the browser as it runs.

Models are delivered on a per request basis. You select a model from the model selector in the workspace. So you can change providers or use local models without changing anything else.

# Bring your own key; requests go straight to the provider
export ANTHROPIC_API_KEY=sk-ant-...
openscience

# Or open a specific project directory
openscience ~/code/my-project

Your keys stay on your device. Sessions, artifacts, and provenance are stored on disk. They can be shared as links.

Four things make working time useful for real work:

• Research ambassadors: A research the agent runs automatically. The expert biology, physicsagain ml agents are also available. Small agents for analyzing and reviewing books and a read-only program mode around it.
• 250+ skills: This includes training (DeepSpeed, PEFT, TRL), testing, dataset work, and cheminformatics. They also cover molecular and clinical biology, papers, LaTeX, mathematics, and cloud computing.
• Scientific databases as tools: UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, and Semantic Scholar are queried. About 30 others are not included.
• A real workplace: Has file tree, editor, terminal, and session history. It provides molecules, structures, genomes, and plots in a row.

Expansion is a first-rate feature. OpenScience supports LSP integration, MCP servers, plugins, and custom agents. It also ships with the TypeScript SDK.

There is an optional managed layer called Atlas. Atlas offers a select set of border models that are billed from a prepaid wallet. It also adds continuous research graph and cloud computing. OpenScience works with Atlas but doesn’t really want it.

OpenScience vs Claude Science

Both tools target the same task. Both run the loop, provide science in line, and prioritize reproducibility. The main difference is the opening and selection of the model.

Claude Science is a refined, independent product with selected integration. OpenScience trades some polish for openness, readability, and provider freedom.

Use Cases with examples

• Machine learning research: An ML developer wants to test the idea of ​​fine tuning. I ml the agent pulls related papers from arXiv, then uses PEFT and TRL capabilities. Writes a training script, runs it, and writes a short report.
• Computational biology: A data scientist studies a protein target. I biology the agent queries UniProt and PDB, and provides an inline structure. It suggests the candidate’s conversion and records the history.
• Cheminformatics: A chemist tests small molecules. Agent queries CheMBL and PubChem for bioactivity data. It applies a filter by code and returns candidates ranked by episodes.
• A budget comparison model: The team does the same work on Claude, then GLM, then local song. Changing a single selection, not rewriting. They compare cost and quality on their own data.

Strengths and Weaknesses

Power:

• It is fully open source under Apache 2.0, so the capabilities and agents are readable and editable.
• Model-agnostic routing eliminates single-vendor lock-in for scientific workflows.
• It works with your infrastructure, so private datasets stay on your systems.
• Extensive tool coverage: 250+ skills and a wealth of scientific information such as tools.
• Extensible with LSP, MCP servers, plugins, and TypeScript SDK.

Weaknesses:

• The agent is not sandboxed; the consent system is not a boundary of isolation.
• You should run it inside a container or VM if you need isolation.
• It’s a small project, so expect rough edges compared to the mature product.
• Turnkey delivery means you manage the provider’s costs and rate limits yourself.
• The quality depends a lot on which model you take each application to.

Interactive Descriptor